Resource pushing method and device, training method and device, electronic equipment and storage medium

Abstract

The invention provides a resource pushing method and device, a training method and device, electronic equipment and a storage medium, belongs to the technical field of artificial intelligence, and particularly relates to the technical fields of resource recommendation, intelligent search, big data and the like. The resource pushing method comprises the following steps: acquiring resource related characteristics for a target object; performing intention detection on the target object based on the resource related features to obtain extended topic features, an extended topic represented by the extended topic features and an initial topic used for the resource related features satisfying a semantic difference degree condition; and determining a target resource from the candidate resources based on the extended topic features, and pushing the target resource to the target object.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technologies, particularly to technologies such as resource recommendation, intelligent search, big data, etc. Specifically, it relates to a resource push method, a training method, an apparatus, an intelligent agent, an electronic device, a storage medium, and a computer program product. Background Art

[0002] With the rapid development of artificial intelligence technologies, users can conveniently browse resource information such as news and videos through terminal devices such as smart phones. Summary of the Invention

[0003] The present disclosure provides a resource push method, a training method, an apparatus, an intelligent agent, an electronic device, a storage medium, and a computer program product.

[0004] According to one aspect of the present disclosure, there is provided a resource push method, including: obtaining resource-related features for a target object; performing intent detection on the target object based on the resource-related features to obtain an extended topic feature, where the extended topic represented by the extended topic feature satisfies a semantic difference degree condition with the initial topic for the resource-related features; determining a target resource from candidate resources based on the extended topic feature, and pushing the target resource to the target object.

[0005] According to another aspect of the present disclosure, there is provided a method for training a deep learning model, including: obtaining sample resource-related features for a sample object; processing the sample resource-related features by a first encoder of the deep learning model to obtain a sample initial topic feature; processing the sample initial topic feature by a second encoder of the deep learning model to obtain a sample extended topic feature; where the sample topic represented by the sample initial topic feature satisfies a semantic similarity condition with the sample initial topic represented by the sample resource-related features; using the sample extended topic feature and the sample initial topic feature as features of a negative sample pair, and training the deep learning model based on a contrast learning mechanism to obtain a trained deep learning model.

[0006] According to another aspect of the present disclosure, there is provided a resource push apparatus, including: a first obtaining module, configured to obtain resource-related features for a target object; an extended topic feature obtaining module, configured to perform intent detection on the target object based on the resource-related features to obtain an extended topic feature, where the extended topic represented by the extended topic feature satisfies a semantic difference degree condition with the initial topic for the resource-related features; a first determining module, configured to determine a target resource from candidate resources based on the extended topic feature, and push the target resource to the target object.

[0007] According to another aspect of the present disclosure, there is provided an apparatus for training a deep learning model, including: a second acquisition module configured to acquire sample resource-related features for a sample object; a sample initial topic feature acquisition module configured to process the sample resource-related features by using a first encoder of the deep learning model to obtain sample initial topic features; a sample extended topic feature acquisition module configured to process the sample initial topic features by using a second encoder of the deep learning model to obtain sample extended topic features; wherein, the sample topic represented by the sample initial topic features and the sample initial topic represented by the sample resource-related features satisfy a semantic similarity condition; a training module configured to use the sample extended topic features and the sample initial topic features as features of a negative sample pair, and train the deep learning model based on a contrastive learning mechanism to obtain a trained deep learning model.

[0008] According to another aspect of the present disclosure, there is provided an agent for artificial intelligence, including: an input module configured to receive input information; a processing module configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method provided in the embodiments of the present disclosure by calling the large model to obtain output information; an output module configured to output the output information obtained by the processing module.

[0009] According to another aspect of the present disclosure, there is provided an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable 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 execute the method provided in the embodiments of the present disclosure.

[0010] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the method provided in the embodiments of the present disclosure.

[0011] According to another aspect of the present disclosure, there is provided a computer program product, including a computer program which, when executed by a processor, implements the method provided in the embodiments of the present disclosure.

[0012] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. Description of the Drawings

[0013] The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. Among them:

[0014] Figure 1 Schematically shows an exemplary system architecture to which the resource push method and apparatus according to the embodiments of the present disclosure can be applied;

[0015] Figure 2 Schematically shows a flowchart of a resource pushing method according to an embodiment of the present disclosure;

[0016] Figure 3 Schematically shows a schematic diagram of the principle of a resource pushing method according to an embodiment of the present disclosure;

[0017] Figure 4 Schematically shows a schematic diagram of the principle of a resource pushing method according to another embodiment of the present disclosure;

[0018] Figure 5 Schematically shows a flowchart of a method for training a deep learning model according to an embodiment of the present disclosure;

[0019] Figure 6 Schematically shows a schematic diagram of the principle of a method for training a deep learning model according to an embodiment of the present disclosure;

[0020] Figure 7 Schematically shows a schematic diagram of a method for training a deep learning model according to another embodiment of the present disclosure;

[0021] Figure 8 Schematically shows a block diagram of a resource pushing device according to an embodiment of the present disclosure;

[0022] Figure 9 Schematically shows a block diagram of a device for training a deep learning model according to an embodiment of the present disclosure;

[0023] Figure 10 Schematically shows a block diagram of the structure of an agent of artificial intelligence according to an embodiment of the present disclosure; and

[0024] Figure 11 Shows a schematic block diagram of an exemplary electronic device that can be used to implement the resource pushing method and the method for training a deep learning model provided by the embodiments of the present disclosure. Detailed Embodiments

[0025] The following describes exemplary embodiments of the present disclosure with reference to the accompanying drawings. Various details of the embodiments of the present disclosure are included to assist in understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0026] In the technical solutions of the present disclosure, the acquisition, storage, and application of the user's personal information involved all comply with the provisions of relevant laws and regulations, take necessary confidentiality measures, and do not violate public order and good customs.

[0027] The inventor has found that the resource information such as videos and news recommended to users has a high similarity with the resources that the users have browsed before. It is easy for users to repeatedly obtain resource information in the same field, resulting in the formation of an "information cocoon" by the users based on the pushed resources. As a result, the pushed resources lack diversity, it is difficult to expand the resources that the users may be interested in, and the user experience is poor.

[0028] Embodiments of the present disclosure provide a resource pushing method, a training method, a device, an intelligent agent, an electronic device, a storage medium, and a computer program product. The resource pushing method includes: obtaining resource-related features for a target object; performing intent detection on the target object based on the resource-related features to obtain extended topic features, where the extended topic represented by the extended topic features satisfies a semantic difference degree condition with the initial topic for the resource-related features; determining a target resource from candidate resources based on the extended topic features, and pushing the target resource to the target object.

[0029] According to the embodiments of the present disclosure, intent detection is performed on the resource-related features to determine the extended topic features for the target object. Since the semantic difference degree between the extended topic features and the initial topic represented by the resource-related features of the target object is relatively large, the extended topic represented by the extended topic features can meet the needs of the target object for exploring new interesting topics. Therefore, by determining the target resource from the candidate resources based on the extended topic features, the target resource can be matched with the extended topic represented by the extended topic features, so as to push resources with different topic types from the resources that have been obtained or browsed by the target object currently, so as to achieve the goal of pushing target resources with diverse topics to the target object, to avoid the target object from obtaining duplicate resources with a high similarity to the resources obtained in the historical period, improve the diversity of resource recommendation, and further improve the user experience.

[0030] Figure 1 An exemplary system architecture to which the resource pushing method and device according to the embodiments of the present disclosure can be applied is schematically shown.

[0031] It should be noted that Figure 1 The illustration is only an example of the system architecture to which the embodiments of the present disclosure can be applied, to help those skilled in the art understand the technical content of the present disclosure, but it does not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios. For example, in another embodiment, the exemplary system architecture to which the resource pushing method and device can be applied may include a terminal device, but the terminal device can implement the resource pushing method and device provided by the embodiments of the present disclosure without interacting with the server.

[0032] Such as Figure 1As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0033] Users can use the terminal devices 101, 102, 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications may be installed on the terminal devices 101, 102, 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only as examples).

[0034] The terminal devices 101, 102, 103 may be various electronic devices with a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop portable computers, and desktop computers, etc.

[0035] The server 105 may be a server that provides various services, such as a background management server that supports the content browsed by users using the terminal devices 101, 102, 103 (only as an example). The background management server may analyze and process data such as received user requests, etc., and feedback the processing results (such as web pages, information, or data, etc. obtained or generated according to user requests) to the terminal devices.

[0036] The server 105 may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system, solving the defects of difficult management and weak business scalability existing in traditional physical hosts and VPS services ("Virtual Private Server", or simply "VPS"). The server 105 may also be a server of a distributed system, or a server combined with a blockchain.

[0037] It should be noted that the resource push method provided by the embodiments of the present disclosure can generally be executed by the terminal devices 101, 102, or 103. Correspondingly, the resource push device provided by the embodiments of the present disclosure can also be set in the terminal devices 101, 102, or 103.

[0038] Alternatively, the resource pushing method provided by the embodiments of the present disclosure can generally also be executed by the server 105. Correspondingly, the resource pushing device provided by the embodiments of the present disclosure can generally be disposed in the server 105. The resource pushing method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103, and / or the server 105. Correspondingly, the resource pushing device provided by the embodiments of the present disclosure can also be disposed in a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103, and / or the server 105.

[0039] For example, when the user is online browsing video resources, the terminal devices 101, 102, 103 can obtain resource-related features related to the user's historical interaction resources, and then send the obtained resource-related features to the server 105. The server 105 performs intent detection on the target content to determine the extended topic features; determines the target resources from the candidate resources according to the extended topic features, and the server 105 pushes the target resources to the terminal devices 101, 102, 103. Or a server or a server cluster capable of communicating with the terminal devices 101, 102, 103, and / or the server 105 obtains resource-related features related to the user's historical interaction resources, and finally realizes the determination of the target resources.

[0040] It should be understood that Figure 1 the numbers of the terminal devices, the network, and the server in

[0041] Figure 2 are merely illustrative. According to the implementation requirements, there can be any number of terminal devices, networks, and servers.

[0042] As Figure 2 shown, the method includes operations S210 to S230.

[0043] In operation S210, resource-related features for the target object are obtained.

[0044] In operation S220, intent detection is performed on the target object based on the resource-related features to obtain extended topic features.

[0045] In operation S230, target resources are determined from the candidate resources based on the extended topic features, and the target resources are pushed to the target object.

[0046] According to an embodiment of the present disclosure, the resource-related feature for a target object may be feature information related to the historical interaction resources of the target object. The historical interaction resources may include resources on which the target object has performed interaction operations during a historical period, such as resources browsed by the target object during a historical period, and resources on which interaction operations such as liking and collection have been performed during a historical period. However, it is not limited thereto. The resource-related feature may also be other types of resources related to the target object, such as resources that have been pushed to the target object during a historical period, resources with the same resource theme as the resources on which the target object has performed interaction operations, and so on. The embodiments of the present disclosure do not limit the specific type of the resource corresponding to the resource-related feature.

[0047] According to an embodiment of the present disclosure, the resource-related feature may include any type of feature information related to the resource, such as the theme information of the resource, the resource content information, the resource comment information, and so on. However, it is not limited thereto. The resource-related feature may also include features of the target object related to the resource, such as the attribute information set by the target object, and so on. The embodiments of the present disclosure do not limit the specific feature type of the resource-related feature, and those skilled in the art may select according to actual needs.

[0048] It should be noted that the acquisition of information involved in the embodiments of the present disclosure, including but not limited to resource-related features, etc., is obtained under the condition of obtaining the authorization of the relevant user or institution, and the purpose of the obtained information is clearly informed as pushing resources of interest to the specified target object before obtaining the authorization, which complies with the provisions of laws and regulations and does not violate public order and good customs.

[0049] According to an embodiment of the present disclosure, performing intention detection on a target object based on resource-related features may include processing the resource-related features based on a trained deep learning model to obtain extended theme features. For example, the deep learning model may be constructed based on any type of deep learning algorithm such as an attention network algorithm, a convolutional neural network algorithm, and so on. The embodiments of the present disclosure do not limit the specific type of constructing the deep learning model.

[0050] According to an embodiment of the present disclosure, performing intention detection on a target object based on resource-related features may also include querying a storage space such as a knowledge graph and a preset mapping table that stores preset extended resource features based on the resource-related features to obtain extended theme features. The embodiments of the present disclosure do not limit the specific manner of performing intention detection on a target object based on resource-related features, as long as it can satisfy performing intention detection based on resource-related features.

[0051] According to an embodiment of the present disclosure, a semantic difference degree condition is satisfied between the extended topic characterizing the extended topic feature and the initial topic for the resource-related feature. The initial topic may include the topic type to which multiple resources corresponding to the resource-related feature belong. The topic type may include the result of classifying the resources. For example, the topic type of video resources may include emotional topic types, teaching topic types, news topic types, and so on. Satisfying the semantic difference degree condition between the extended topic and the initial topic may indicate a low semantic relevance between the extended topic and the initial topic. Therefore, determining the target resource from the candidate resources based on the extended topic feature of the extended topic can make the target resource close to the topic type indicated by the extended topic. Thus, by pushing the target resource to the target object, the resource range obtained by the target object can be extended, the interest points of the target object can be extended, the information cocoon effect caused by the target object obtaining resources with high similarity can be avoided, and the diversity of resource pushing can be improved.

[0052] It should be understood that the topic type of the same resource may include one or more. Or the same topic type may also include multiple topic words. For example, the topic type of video resources may have topic words such as "lecture", "artificial intelligence", and "animation".

[0053] According to an embodiment of the present disclosure, the resource-related feature includes at least one of the following: the historical resource feature for the historical resource; the object attribute feature related to the target object.

[0054] According to an embodiment of the present disclosure, the historical resource is a resource for which the target object has performed an interaction operation. For example, the historical resource may be a resource for which the target object has performed interaction operations such as clicking, liking, commenting, and collecting. The resource feature for the historical resource may include any feature information related to the historical resource, such as the topic of the historical resource, the resource content, the resource comment, etc. However, it is not limited to this. The historical resource feature may also include feature information related to the resource attribute of the historical resource, such as the resource text length and the resource playback length of the historical resource. The specific type of feature information of the historical resource feature in the embodiments of the present disclosure is not limited, as long as it is related to the historical resource.

[0055] According to an embodiment of the present disclosure, the object attribute feature related to the target object may represent feature information related to the target object, such as the identity of the target object, the type of interaction operation, and the resource preference. The specific type of the object attribute feature in the embodiments of the present disclosure is not limited, as long as it is related to the target object.

[0056] It should be noted that the acquisition of information involved in the embodiments of the present disclosure, including but not limited to historical resource characteristics, object attribute characteristics, etc., is obtained under the condition of obtaining the authorization of relevant users or institutions, and the purpose of the obtained information is clearly informed as pushing resources of interest to the specified target object before obtaining the authorization, which complies with the provisions of laws and regulations and does not violate public order and good customs.

[0057] In one example, the resource-related characteristics are related to the satisfactory interaction resources of the target object in the historical period. The satisfactory interaction resources can represent the historical resources that the target object is more interested in among the historical resources for which interaction operations have been performed. For example, the satisfactory interaction resources can be historical resources with a longer browsing duration by the target object. For another example, the satisfactory interaction resources can also be historical resources that the target object has liked or commented on. The embodiments of the present disclosure do not limit the specific determination method of the satisfactory interaction resources, as long as they can represent the interest demand of the target object for the satisfactory interaction resources in the historical period.

[0058] In one example, the satisfactory interaction resources can be determined based on the following formula (1).

[0059] (1);

[0060] In formula (1), sat represents the satisfactory interaction resources, sat_t represents the satisfactory consumption rate, play_time represents the browsing duration of the target object for the historical resources, or play_time represents the playback duration of the historical resources for the target object, s_time represents the preset browsing duration of the historical resources, or the resource playback duration of the historical resources; s_like represents that the target object has performed a "like" interaction operation on the historical resources, s_share represents that the target object has performed a "share" interaction operation on the historical resources, a represents the preset threshold, and a can be a parameter greater than 0 and less than 1.

[0061] In this example, when the historical interaction resources meet at least one of the following preset conditions, it can be determined that the historical resources are satisfactory interaction resources. The preset conditions may include: the satisfactory consumption rate is greater than 1; the ratio between play_time and s_time is greater than the preset threshold; the target object has performed a "share" interaction operation on the historical interaction resources; the target object has performed a "like" interaction operation on the historical interaction resources.

[0062] Among them, sat_t representing the satisfactory consumption rate can be determined based on a preset formula. For example, the ratio between play_time and s_time can be corrected by a preset correction parameter to obtain the satisfactory consumption rate.

[0063] For another example, the satisfactory consumption rate can be determined based on the following formula (2)

[0064] (2);

[0065] Among them, b, c, and d represent preset empirical parameters.

[0066] According to an embodiment of the present disclosure, by performing intention detection on a target object based on resource-related features of a satisfactory interaction resource for the target object, it is possible to make the semantic difference between the extended theme feature and the initial theme not too large when the extended theme feature obtained has a difference in semantics from the initial theme. Thus, the target resource determined based on the extended theme feature can meet the exploration needs of the target object for new types of resources, improving the quality of resource push.

[0067] In one example, the object attribute feature includes n types of object attribute sub-features, and the object attribute feature matrix representing the object attribute feature can be expressed as fea_user = R c × k . The object attribute feature matrix can be a matrix of c×k dimensions. When the embedding dimension of the historical resource feature is k and the number of historical resource features is x, the historical resource feature can be represented based on a matrix of dimensions x×k. At the same time, a prompt token can be set as prompt information. The historical resource feature can represent t initial themes of interest to the target object, and the prompt token can be represented as a matrix of dimensions t×k. Intention detection is performed on the target object based on the prompt information, historical resource feature, and object attribute feature to obtain an extended theme feature. For example, the concatenation result of the prompt token, historical resource feature matrix, and object attribute feature matrix can be input into the first detection network to perform intention detection on the target object and output an extended theme matrix. The matrix input into the first detection network can be expressed as fea = R (t+x+c) × k .

[0068] In one example, the extended theme feature output by the first detection network can represent t extended themes, and the extended theme feature can be expressed as a matrix of t×k dimensions.

[0069] According to an embodiment of the present disclosure, performing intention detection on a target object based on resource-related features to obtain an extended theme feature may include: performing feature fusion on the resource-related features based on an attention mechanism to obtain an initial theme feature representing the initial theme; performing intention detection based on the initial theme feature to obtain an extended theme feature.

[0070] According to an embodiment of the present disclosure, feature fusion of resource-related features based on an attention mechanism may include performing self-attention fusion on the resource-related features based on a multi-head attention mechanism to obtain an initial topic feature. The initial topic feature may represent multiple initial topics, and based on prompt information representing the multiple initial topics as positional embedding encoding, it may be concatenated with the resource-related features, and the concatenation result of the resource-related features and the multiple prompt information may be processed based on a self-attention mechanism to obtain an initial topic feature representing the multiple initial topics.

[0071] According to an embodiment of the present disclosure, intent detection based on the initial topic feature may include processing the initial topic feature based on an attention network algorithm to implement attention fusion of one or more initial topics represented by the initial topic feature and resource-related features for a target object to detect an extended topic potentially of interest to the target object, so that the extended topic feature can represent multiple extended topics potentially of interest to the target object.

[0072] According to an embodiment of the present disclosure, the target resource is determined from candidate resources based on the extended topic feature and the initial topic feature. For example, an extended topic feature and an initial topic feature may be processed by a trained neural network model to obtain the semantics of the resource-related features in fusion, as well as a fused topic feature of the semantics of the initial topic and the extended topic. By matching the fused topic feature with a preset candidate resource feature library, a target resource feature matching the fused topic feature is obtained, and the candidate resource corresponding to the target resource feature is determined as the target resource.

[0073] In one example, a candidate resource feature representing the candidate resource, as well as the extended topic feature and the initial topic feature, may also be processed based on a trained neural network layer to obtain a demand weight for the candidate resource. Thus, the target resource may be determined from multiple candidate resources through the demand weights corresponding to the multiple candidate resources.

[0074] According to an embodiment of the present disclosure, by obtaining resource-related features and performing attention fusion on the resource-related features based on the attention mechanism, the initial topic features can implicitly represent the distribution of multiple initial topic interest points of the target object under the condition that the resource information of multiple historical resources interacted with the target object can be fused. By performing intention detection on the target object based on the initial topic features to obtain extended topic features, when the extended topic represented by the extended topic features satisfies the semantic difference condition with the initial topic, the semantic difference between the extended topic semantics represented by the extended topic features and the initial topic semantics will not have too large a deviation. Thus, based on the extended topic features, the initial topic features, and the candidate resource features, the demand weight of the candidate resource can be determined, so that the demand weight can more accurately represent the difference between the candidate resource and the initial topic currently interested by the target object, as well as the difference between the candidate resource and the extended topic potentially interested by the target object. Therefore, based on the demand weight, the target resource that meets the extended interest demand of the target object can be determined from the candidate resources, so that the target resource not only represents the potential new interest points of the target object, but also does not have too large a difference from the topic currently interested by the target object, thereby realizing the resource information that can meet the breaking of the "information cocoon" to be pushed to the target object to improve the user experience of the target object.

[0075] In one example, the resource-related features and the candidate resource features can be processed based on a trained deep learning model to obtain the demand weight of the candidate resource. Thus, based on the demand weight corresponding to each of one or more candidate resources, the target resource that can represent the extended topic interest of the target object can be determined from the candidate resources. The resource pushing method provided by the embodiments of the present disclosure will be explained in detail below based on Figure 3 the embodiments shown.

[0076] Figure 3 Schematically shows a schematic diagram of the principle of the resource pushing method according to an embodiment of the present disclosure.

[0077] As Figure 3 shown, the deep learning model may include a first encoder M311, a second encoder M312, and a resource evaluation network M320. The first encoder M311 and the second encoder M312 may be neural network layers constructed based on the multi-head attention mechanism. For example, each of the first encoder M311 and the second encoder M312 may have a multi-head attention layer, a feed forward layer, and two residual normalization (Add&Norm) layers. The resource evaluation network M320 may be constructed based on any neural network algorithm. For example, the resource evaluation network M320 may be constructed based on the multi-layer perceptron (MLP) algorithm.

[0078] Resource-related features may include object attribute features 320 for a target object and multiple satisfactory resource features 310. The object attribute features 320 and the multiple satisfactory resource features 310 are input into the first encoder M311, and initial topic features representing multiple initial topics are output. The initial topic features are input into the second encoder M312, and extended topic features are output. The extended topics represented by the extended topic features satisfy the semantic difference degree condition with the initial topics represented by the initial topic features. The candidate resource features 330 may be determined by performing semantic feature extraction on candidate resources based on a preset feature extraction network. Thus, by inputting the extended topic features, the initial topic features, and the candidate resource features 330 into the resource evaluation network M320, the demand weights for the candidate resources can be obtained. The multiple candidate resources are sorted according to the demand weights corresponding to each of the multiple candidate resources to obtain multiple candidate resources arranged in order. The top n candidate resources in the ranking are used as target resources representing potential extended topics of the target object. By pushing the target resources to the target object, the target resources for satisfying the user's expanded interest point range are obtained, and the diverse needs of the user for resource information are satisfied.

[0079] In one example, demand intention evaluation is performed based on the extended topic features, the initial topic features, the candidate resource features, and the object attribute features to obtain the demand weights for the candidate resources. For example, a resource evaluation network constructed based on a multi-layer perceptron can be used to process the extended topic features, the initial topic features, the candidate resource features, and the object attribute features to obtain the demand weights for the candidate resources. Thus, in the case of further integrating the attribute information of the target object, the demand weights can further represent the matching degree between the candidate resources and the object attributes of the target object, so that the target resources determined based on the demand weights can represent extended topics that satisfy the semantic difference degree condition with the semantics of the initial topics, and the semantic difference from the initial topics is not too large, and at the same time, they can also match the object attributes of the target object, thereby further enhancing the demand for the target resources to expand the resource interest points of the target object and improving the accuracy of resource push.

[0080] In one example, demand intention evaluation of candidate resources based on the extended topic features and the candidate resource features for the candidate resources includes: performing demand intention evaluation on the candidate resources based on the extended topic features, the candidate resource features, and the interaction scenario features in the resource-related features to obtain initial demand weights. For example, a trained resource evaluation network can be used to process the extended topic features, the initial topic features, the candidate resource features, and the interaction scenario features to obtain the initial demand weights.

[0081] In one example, the demand intention evaluation of candidate resources based on the extended theme features and candidate resource features for the candidate resources may include: performing demand intention evaluation based on the extended theme features, initial theme features, candidate resource features, and object attribute features to obtain an initial demand weight. For example, a trained resource evaluation network may be used to process the extended theme features, initial theme features, candidate resource features, and object attribute features to obtain the initial demand weight.

[0082] In one example, the demand intention evaluation of candidate resources based on the extended theme features and candidate resource features for the candidate resources may include: performing demand intention evaluation based on the extended theme features, initial theme features, candidate resource features, interaction scenario features, and object attribute features to obtain an initial demand weight. For example, a trained resource evaluation network may be used to process the extended theme features, initial theme features, candidate resource features, interaction scenario features, and object attribute features to obtain the initial demand weight.

[0083] According to an embodiment of the present disclosure, the resource-related features may further include interaction scenario features, and the interaction scenario features may represent the resource interaction scenario of the target object during a specified period. For example, the interaction scenario features may represent attribute information related to interaction operation behaviors such as page refresh operations and comment operations of the target object during a specified duration before the current moment.

[0084] In one example, demand intention evaluation is performed based on the extended theme features, initial theme features, candidate resource features, object attribute features, and interaction scenario features to obtain the demand weight of the candidate resources. For example, a resource evaluation network constructed based on a multi-layer perceptron may be used to process the extended theme features, initial theme features, candidate resource features, object attribute features, and interaction scenario features to obtain the demand weight of the candidate resources. Thereby, the demand weight can more fully represent the semantic similarity between the resource information of the candidate resources and the extended theme, as well as the semantic difference degree with the initial theme; at the same time, the demand weight can also represent the matching degree between the candidate resources and the interaction operation behaviors and object attribute information of the target object, so that the target resources that match the interest expansion intention and actual needs of the target object can be more accurately determined from the candidate resources based on the demand weight, and the resource exploration needs of the target object can be satisfied by pushing the target resources to the target object.

[0085] According to an embodiment of the present disclosure, determining a target resource from candidate resources based on extended theme features includes: performing demand intention evaluation on the candidate resources based on the extended theme features and candidate resource features for the candidate resources to obtain an initial demand weight; updating the initial demand weight based on the similarity between the extended theme features and the candidate resource features to obtain the demand weight for the candidate resources; and determining the target resource from at least one candidate resource based on the demand weight.

[0086] According to an embodiment of the present disclosure, the demand intention evaluation of candidate resources based on the extended theme features and candidate resource features for the candidate resources may include processing the extended theme features and the candidate resource features by using a trained resource evaluation network to obtain an initial demand weight.

[0087] According to an embodiment of the present disclosure, the initial demand weight is updated based on the similarity between the extended theme features and the candidate resource features. It may include updating the initial demand weight by using the similarity between the extended theme features and the candidate resource features as a weight correction parameter to obtain an updated demand weight. The similarity between the extended theme features and the candidate resource features may represent the degree of correlation between the resource information of the candidate resource and the semantics of the extended theme. By updating the initial weight with the weight correction parameter determined based on the similarity between the extended theme features and the candidate resource features, the updated demand weight can more accurately represent the degree of correlation between the candidate resource and the semantics of the extended theme, so that the target resource determined from the candidate resources based on the demand weight can better match the extended theme, further accurately meet the diverse needs of the target object to expand the interest points by obtaining the target resource, and improve the diversity and accuracy of the target resource push.

[0088] In one example, the initial demand weight may be updated based on the following formula (3) to obtain an updated demand weight.

[0089] (3);

[0090] where exp() represents the exponential function with the base e of the natural logarithm. Y represents the updated demand weight, y pr is the initial demand weight output by the resource evaluation network, cos_sim represents the similarity between the extended theme features and the candidate resource features, bias and new_bias are hyperparameters, and new_bias represents the exploration intensity of the semantic difference between the extended theme and the initial theme. The degree of correlation between the candidate resource characterized by the demand weight and the extended theme can be adjusted by setting the specific values of the hyperparameters bias and new_bias, so as to realize the determination of a target resource that better adapts to the extended theme requirements of the target object.

[0091] In one example, the hyperparameters bias and new_bias may be model parameters of a deep learning model, and the hyperparameters bias and new_bias may be adjusted by training the deep learning model.

[0092] It should be noted that the resource push method provided in the embodiments of the present disclosure can be executed based on a trained deep learning model, and the deep learning model can include a first detection network and a resource evaluation network. The hyperparameters bias and new_bias can be the model parameters of the resource evaluation network or the model parameters of the output layer cascaded with the resource evaluation network, and the embodiments of the present disclosure do not limit this.

[0093] Figure 4 Schematically shows a schematic diagram of the principle of the resource push method according to another embodiment of the present disclosure.

[0094] As Figure 4 shown, the deep learning model can include a feature detection network M410 and a first resource evaluation network M420. The feature detection network M410 can include a feature encoder M411 and an inference encoder M412. It should be noted that in this embodiment, the first detection network is the feature detection network M410, the first encoder is the feature encoder M411, the second encoder is the inference encoder M412, and the first resource evaluation network M420 is the resource evaluation network. The feature encoder M411 and the inference encoder M412 can each be constructed based on the multi-head attention algorithm, and the resource evaluation network can be constructed based on the multi-layer perceptron algorithm and the pooling algorithm.

[0095] As Figure 4 shown, the resource-related features can include multiple satisfactory resource features, object attribute features, and interaction scenario features. The multiple satisfactory resource features can each characterize different satisfactory interaction resources of the target object. The prompt information can be a prompt token, and the prompt information is used as a positional encoding to represent t different initial topics. For example, based on the role of the prompt token as the [CLS] token in the BERT (Bidirectional Encoder Representations from Transformers) model, the multiple sub-features in the input resource-related features can be aligned with the positions corresponding to the t initial topics, so as to realize the initial topic feature fusing the feature information such as the resource information and the object attribute information of the resource-related features. Among them, t is an integer greater than 1. Input the concatenation result of the multiple satisfactory resource features, object attribute features, and prompt information into the feature encoder M411, and output the initial topic features representing the t initial topics. The initial topic features can fuse the resource information of multiple satisfactory interaction resources and the object attribute information of the target object. Input the initial topic features into the inference encoder M412, and output the extended topic features representing multiple extended topics. The extended topic features can represent multiple extended topics under the condition of fully fusing the resource information of multiple satisfactory interaction resources and the object attribute information of the target object.

[0096] Input the concatenation result of candidate features, initial topic features, extended topic features, object attribute features, and interaction scenario features for candidate resources into the first resource evaluation network M420 to output the initial weight for the candidate resources. The first resource evaluation network M420 can pool the extended topic features and the initial topic features into a matrix of 1×x×k dimensions based on the average pooling algorithm. Process the candidate feature object attribute features and interaction scenario features, as well as the matrix of 1×x×k dimensions, through the multi-layer perceptron algorithm to output the initial requirement weight. Update the initial requirement weight with a correction parameter determined based on the similarity between the extended topic features and the candidate resource features to determine the requirement weight for the candidate resources. Sort the multiple candidate resources based on the respective requirement weights of the multiple candidate resources, and determine the target resource based on the sorting result.

[0097] Based on the resource push method provided in the above embodiments, an embodiment of the present disclosure further provides a method for training a deep learning model. The trained deep learning model determined by this method for training a deep learning model can be used in the above resource push method.

[0098] Figure 5 Schematically shows a flowchart of a method for training a deep learning model according to an embodiment of the present disclosure.

[0099] As Figure 5 shown, the method for training a deep learning model includes operations S510 to S540.

[0100] In operation S510, obtain sample resource-related features for a sample object.

[0101] In operation S520, process the sample resource-related features using the first encoder of the deep learning model to obtain sample initial topic features;

[0102] In operation S530, process the sample initial topic features using the second encoder of the deep learning model to obtain sample extended topic features; wherein, the sample topic represented by the sample initial topic features and the sample initial topic represented by the sample resource-related features satisfy the semantic similarity condition;

[0103] In operation S540, use the sample extended topic features and the sample initial topic features as the features of a negative sample pair, and train the deep learning model based on the contrast learning mechanism to obtain the trained deep learning model.

[0104] The technical terms involved in the method for training a deep learning model provided by an embodiment of the present disclosure, including but not limited to sample objects, sample resource-related features, sample extended topic features, etc., have the same or corresponding attributes as the technical terms involved in the resource pushing method provided by the embodiment of the present disclosure, including but not limited to target objects, resource-related features, extended topic features, etc. The embodiments of the present disclosure will not elaborate herein.

[0105] According to an embodiment of the present disclosure, the sample theme characterized by the sample initial theme feature and the sample initial theme characterized by the sample resource-related feature satisfy the semantic similarity condition. For example, the semantic similarity between the sample theme and the sample initial theme satisfies a preset semantic similarity threshold.

[0106] According to an embodiment of the present disclosure, by using the sample extended topic feature and the sample initial theme feature as the features of the negative sample pair to implement the training of the deep learning model based on the contrast learning mechanism, it can make the semantic difference degree between the sample extended topic feature output by the second encoder and the sample initial theme feature output by the first encoder gradually increase during the training process of the deep learning model until the training condition is satisfied. Since the semantic similarity between the sample theme characterized by the sample initial theme feature and the sample initial theme characterized by the sample resource-related feature is relatively high, the extended topic feature output by the first encoder and the second encoder in the first detection network of the trained deep learning model by processing the resource-related features can make the semantic difference degree between the extended topic characterized by the extended topic feature and the initial theme characterized by the resource-related feature satisfy the semantic difference degree condition. Thus, the deep learning model can obtain the extended topic feature representing the extended topic by fusing the resource-related features representing the initial theme, and further enable the target resource determined based on the extended topic feature to meet the user's interest point expansion requirement, realizing diversified resource pushing for the target object.

[0107] It should be noted that the acquisition of information involved in the method for training a deep learning model provided by the embodiment of the present disclosure, including but not limited to information such as sample resource-related features, is obtained under the condition of obtaining the authorization of the relevant user or institution, and the purpose of the obtained information is clearly informed as pushing resources of interest to the specified target object before obtaining the authorization, which complies with the provisions of laws and regulations and does not violate public order and good customs.

[0108] According to an embodiment of the present disclosure, the deep learning model further includes a second detection network. The second detection network can be constructed based on a decoder using an attention algorithm.

[0109] According to an embodiment of the present disclosure, training a deep learning model based on a contrastive learning mechanism includes: processing the extended topic features by a second detection network of the deep learning model to obtain sample topic features; using the sample extended topic features and the sample initial topic features as the features of a negative sample pair, and using the sample topic features and the sample initial topic features as the features of a positive sample pair, and training the deep learning model based on the contrastive learning mechanism.

[0110] According to an embodiment of the present disclosure, by using the sample extended topic features and the sample initial topic features as the feature representations of a negative sample pair, and using the sample topic features and the sample initial topic features as the feature representations of a positive sample pair, during the training process of the deep learning model, the semantic difference degree between the sample extended topic features output by the second encoder and the sample initial topic features can be gradually enlarged, and the semantic difference degree between the sample initial topic features output by the first encoder and the sample topic features output by the second detection network can be gradually reduced. Thus, the first detection network of the deep learning model can more accurately detect the resource information and object attribute information that have a certain correlation with the feature representation related to the sample resource, and satisfy the semantic difference degree condition with the initial topic, and the sample extended topic features; the second detection network can also restore the sample topic features representing the sample initial topic based on the sample extended topic features. Thus, the first detection network of the trained deep learning model can process the features related to the sample resource to detect the extended topic features that satisfy the semantic difference degree condition with the initial topic, and the semantic difference degree with the initial topic is not too large, and at the same time fully integrate the resource information and object attribute information. Furthermore, the target resource that meets the needs of the target object's interest topic expansion can be determined from the candidate resources through the extended topic features, improving the resource exploration function for the target object, meeting the actual needs of the target object to diversely obtain resource information, and breaking through the "information cocoon" limitation for the target object.

[0111] Figure 6 Schematically shows a schematic diagram of the principle of a method for training a deep learning model according to an embodiment of the present disclosure.

[0112] As Figure 6As shown, the deep learning model may include a first detection network M610 and a second detection network M620. The first detection network M610 includes a first encoder M611 and a second encoder M612. The second detection network M620 may be constructed based on a decoder of an attention algorithm. The sample resource-related features may include multiple sample satisfaction resource features and sample object attribute features. The prompt information may be used as positional encoding and input, together with the multiple sample satisfaction resource features and sample object attribute features, into the first encoder M611 to output sample initial topic features. The sample initial topic features are input into the second encoder M612 to output sample extended topic features. The sample extended topic features are input into the second detection network M620 to output sample topic features. Taking the sample topic features and the sample initial topic features as the feature representations of positive samples, and taking the sample initial topic features and the sample extended topic features as the feature representations of negative samples to implement training of the deep learning model based on a contrast learning mechanism can increase the semantic difference degree between the initial topic represented by the sample initial topic features and the sample extended topic represented by the sample extended topic features during the training process of the deep learning model, and reduce the semantic difference degree between the sample topic features and the sample initial features. Thereby, the first encoder in the trained deep learning model can more fully integrate the resource information of the sample satisfaction resources represented by the sample resource-related features and the object attribute information represented by the object attribute features, and enable the sample initial topic features to more accurately represent the sample initial topic. At the same time, the second encoder can detect, based on the sample initial topic features, the sample extended topic features corresponding to the potential extended interest topics of the target object, and can restore, based on the second detection network, the sample topic features representing the initial topic, so that the sample extended topic features output by the second encoder can meet the extended topic requirements with a relatively large semantic difference degree from the initial topic, and can also make the semantic of the sample extended topic features and the initial topic representation have a certain degree of semantic correlation. Furthermore, the target resources determined according to the sample extended topic features can meet the interest point extension requirements of the target object, improving the diversity and accuracy of target resource push.

[0113] According to an embodiment of the present disclosure, training the deep learning model based on a contrast learning mechanism may further include: determining first difference information between the extended topic features and the sample topic features; and training the deep learning model based on the first difference information and second difference information characterizing the difference between the sample topic features and the sample initial topic features.

[0114] According to an embodiment of the present disclosure, the first difference information or the second difference information may be represented based on cosine similarity, but is not limited thereto. The first difference information or the second difference information may also be represented based on other data indicating the degree of difference, such as Euclidean distance. The embodiment of the present disclosure does not limit the specific representation manner of the first difference information or the second difference information.

[0115] According to an embodiment of the present disclosure, the first difference information and the second difference information can be processed based on a loss function for a contrastive learning mechanism to obtain a feature loss value, and the model parameters of the deep learning model can be adjusted based on the loss value until the training conditions are met, thereby obtaining a trained deep learning model.

[0116] In one example, the first difference information and the second difference information can be processed based on a BPR loss function (Bayesian Personalized Ranking Loss) to obtain a feature loss value.

[0117] For example, the feature loss value can be determined based on the following formula (4).

[0118] (4);

[0119] Wherein, σ represents the sigmoid activation function, sim() represents the cosine similarity function, interest_hid represents the initial topic feature of the sample, interset_recover represents the topic feature of the sample output by the second detection network, and interest_new represents the extended topic feature of the sample output by the second encoder.

[0120] In one example, the deep learning model may further include a resource evaluation network, which is configured to process the initial topic feature of the sample, the extended topic feature of the sample, and the candidate resource feature of the sample to obtain a sample demand weight. The sample demand weight and the label demand weight for the candidate resource feature of the sample can be processed through a loss function to obtain a weight loss value, and the deep learning model can be trained based on the weight loss value to obtain a trained deep learning model.

[0121] For example, the first encoder and the second encoder of the deep learning model are obtained after being trained based on the feature loss value. The model parameters of the resource evaluation network can be adjusted based on the weight loss value until the training conditions are met, thereby obtaining a trained deep learning model.

[0122] Exemplarily, the weight loss value can be determined based on the following formula (5).

[0123] (5);

[0124] Wherein, y represents the label demand weight, is the sample demand weight output by the resource evaluation network, and Loss is the weight loss value.

[0125] Figure 7 Schematically shows a schematic diagram of a method for training a deep learning model according to another embodiment of the present disclosure.

[0126] AsFigure 7 As shown, the deep learning model may include a first encoder, a second encoder, and a second detection network, where the first encoder and the second encoder may form a first detection network. The sample resource-related features may include multiple sample satisfaction resource features, sample object attribute features, and sample interaction scenario features. The prompt information may be used as positional encoding and input into the first encoder together with the multiple sample satisfaction resource features and sample object attribute features, and the sample initial topic features representing the initial topic of the sample are output. The sample initial topic features are input into the second encoder, and the sample extended topic features representing the extended topic of the sample are output. The sample extended topic features are input into the second detection network, and the sample topic features capable of restoring the semantic attributes of the sample initial topic are output. The sample initial topic features and the sample extended topic features are used as the feature representations of the negative sample pair, and the sample initial topic features and the sample topic features are used as the feature representations of the positive sample pair. Based on the contrast learning mechanism, the first encoder, the second encoder, and the second detection network of the deep learning model are trained to obtain the trained first encoder, second encoder, and second detection network.

[0127] After obtaining the trained first encoder, second encoder, and second detection network, the sample extended topic features, the sample initial topic features, the sample candidate resource features corresponding to the sample candidate resources, the sample object attribute features, and the sample interaction scenario features are input into the resource evaluation network, and the sample demand weights are output. Based on the weight difference information between the sample demand weights and the label demand weights for the sample candidate resources, the resource evaluation network is trained, and then the trained resource evaluation network is obtained. Based on the trained first encoder and second encoder, and the trained resource evaluation network, the trained deep learning model for the resource push method provided in the embodiments of the present disclosure can be determined.

[0128] Figure 8 A block diagram of a resource push device according to an embodiment of the present disclosure is schematically shown.

[0129] As Figure 8 shown, the resource push device 800 includes: a first acquisition module 810, an extended topic feature acquisition module 820, and a first determination module 830.

[0130] The first acquisition module 810 is configured to acquire resource-related features for a target object.

[0131] The extended topic feature acquisition module 820 is configured to perform intent detection on the target object based on the resource-related features to obtain extended topic features, where the extended topic represented by the extended topic features satisfies a semantic difference degree condition with the initial topic for the resource-related features.

[0132] The first determination module 830 is configured to determine a target resource from candidate resources based on extended topic features, and push the target resource to a target object.

[0133] According to an embodiment of the present disclosure, the first determination module 830 includes: an initial requirement weight obtaining unit, a requirement weight obtaining unit, and a target resource determination unit.

[0134] The initial requirement weight obtaining unit is configured to perform a requirement intention evaluation on candidate resources based on extended topic features and candidate resource features for the candidate resources, to obtain an initial requirement weight.

[0135] The requirement weight obtaining unit is configured to update the initial requirement weight based on the similarity between the extended topic features and the candidate resource features, to obtain a requirement weight for the candidate resources.

[0136] The target resource determination unit is configured to determine a target resource from at least one candidate resource based on the requirement weight.

[0137] According to an embodiment of the present disclosure, the initial requirement weight obtaining unit includes a requirement intention evaluation subunit.

[0138] The requirement intention evaluation subunit is configured to perform a requirement intention evaluation on candidate resources based on interaction scenario features among the extended topic features, the candidate resource features, and resource-related features; wherein the interaction scenario features characterize the resource interaction scenario of the target object in a specified period.

[0139] According to an embodiment of the present disclosure, the extended topic feature obtaining module 820 includes: an initial topic feature obtaining unit and an extended topic feature obtaining unit.

[0140] The initial topic feature obtaining unit is configured to perform feature fusion on resource-related features based on an attention mechanism, to obtain initial topic features representing an initial topic.

[0141] The extended topic feature obtaining unit is configured to perform intent detection based on the initial topic features, to obtain extended topic features; wherein the target resource is determined from candidate resources based on the extended topic features and the initial topic features.

[0142] According to an embodiment of the present disclosure, the resource-related features include at least one of the following: historical resource features for historical resources, where the historical resources are resources on which the target object has performed interaction operations; object attribute features related to the target object.

[0143] Figure 9 Schematically shows a block diagram of an apparatus for training a deep learning model according to an embodiment of the present disclosure.

[0144] As Figure 9As shown, the apparatus 900 for training a deep learning model includes: a second acquisition module 910, a sample initial topic feature acquisition module 920, a sample extended topic feature acquisition module 930, and a training module 940.

[0145] The second acquisition module 910 is configured to acquire sample resource-related features for a sample object.

[0146] The sample initial topic feature acquisition module 920 is configured to process the sample resource-related features by using a first encoder of the deep learning model to obtain sample initial topic features.

[0147] The sample extended topic feature acquisition module 930 is configured to process the sample initial topic features by using a second encoder of the deep learning model to obtain sample extended topic features; wherein, the sample topic represented by the sample initial topic features and the sample initial topic represented by the sample resource-related features satisfy a semantic similarity condition.

[0148] The training module 940 is configured to use the sample extended topic features and the sample initial topic features as the features of a negative sample pair, and train the deep learning model based on a contrastive learning mechanism to obtain a trained deep learning model.

[0149] According to an embodiment of the present disclosure, the deep learning model further includes a second detection network. The training module 940 includes: a sample topic feature acquisition unit and a training unit.

[0150] The sample topic feature acquisition unit is configured to process the extended topic features by using the second detection network of the deep learning model to obtain sample topic features.

[0151] The training unit is configured to use the sample extended topic features and the sample initial topic features as the features of a negative sample pair, and use the sample topic features and the sample initial topic features as the features of a positive sample pair, and train the deep learning model based on a contrastive learning mechanism.

[0152] According to an embodiment of the present disclosure, the training unit includes: a first difference information determination subunit and a training subunit.

[0153] The first difference information determination subunit is configured to determine first difference information between the extended topic features and the sample initial topic features.

[0154] The training subunit is configured to train the deep learning model based on the first difference information and second difference information representing the difference between the sample topic features and the sample initial topic features.

[0155] Figure 10 Schematically shows a structural block diagram of an intelligent agent of artificial intelligence according to an embodiment of the present disclosure.

[0156] In an embodiment of the present disclosure, as Figure 10 shown, the AI agent 1000 may include an input module 1010, a processing module 1020, and an output module 1030.

[0157] The input module 1010 is configured to receive input information;

[0158] The processing module 1020 is configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, execute the resource push method provided according to the embodiment of the present disclosure by invoking the large model, or execute the method for training a deep learning model provided according to the embodiment of the present disclosure by invoking the large model, and obtain output information;

[0159] The output module 1030 is configured to output the output information obtained by the processing module.

[0160] According to an embodiment of the present disclosure, the input module 1010 is responsible for receiving or sensing information such as queries, requests, instructions, signals, or data from the outside world (such as users or the external environment), and converting it into a format that the AI agent 1000 can understand and process. The input module 1010 is the primary link for the AI agent 1000 to interact with the outside world, enabling the AI agent 1000 to efficiently and accurately obtain necessary "sensory" information from the outside world and respond to this information.

[0161] In an example, the input module 1010 may input the resource-related features described above or sample resource-related features, etc.

[0162] In an example, the processing module 1020 is the core support for the AI agent 1000 to handle complex tasks. The processing module 1020 may execute the resource push method and the method for training a deep learning model described above.

[0163] In an example, the performance of the processing module 1020 may be closely related to the large model on which the AI agent 1000 is based. To fully utilize the capabilities of the large model, the internal structure of the processing module 1020 can be designed to be highly configurable and extensible to cope with various different types of tasks and requirements in real scenarios.

[0164] In an example, after the AI agent 1000 obtains the resource-related features, the processing module 1020 may use the large model to process the resource-related features to obtain a target resource, and transfer the target resource to the output module 1030.

[0165] It can be understood that although large models have excellent language understanding and generation capabilities, like humans, they can perform very limited tasks without any tools. After the AI agent 1000 is given the ability to call tools, tasks such as performing mathematical operations with the help of a calculator, performing data analysis with the help of Python, and obtaining weather forecasts with the help of a search engine can be achieved.

[0166] In the example, the output module 1030 can output the target resources or the trained deep learning model described above.

[0167] According to an embodiment of the present disclosure, the AI agent 1000 can simply and effectively improve the degree of intelligence, and improve flexibility and versatility.

[0168] 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.

[0169] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable 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 execute the method as described above.

[0170] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the method as described above.

[0171] According to an embodiment of the present disclosure, a computer program product includes a computer program, and the computer program implements the method as described above when executed by a processor.

[0172] Figure 11 FIG. shows a schematic block diagram of an exemplary electronic device that can be used to implement the resource pushing method and the method for training a deep learning model provided by the embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0173] As Figure 11As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded from a storage unit 1108 into a random access memory (RAM) 1103. In the RAM 1103, various programs and data required for the operation of the device 1100 can also be stored. The computing unit 1101, the ROM 1102, and the RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0174] Multiple components in the device 1100 are connected to the I / O interface 1105, including: an input unit 1106, such as a keyboard, a mouse, etc.; an output unit 1107, such as various types of displays, speakers, etc.; a storage unit 1108, such as a magnetic disk, an optical disc, etc.; and a communication unit 1109, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1109 allows the device 1100 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0175] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1101 executes the various methods and processes described above, such as the resource pushing method, the method of training a deep learning model. For example, in some embodiments, the resource pushing method, the method of training a deep learning model can be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as the storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 1100 via the ROM 1102 and / or the communication unit 1109. When the computer program is loaded into the RAM 1103 and executed by the computing unit 1101, one or more steps of the resource pushing method, the method of training a deep learning model described above can be executed. Alternatively, in other embodiments, the computing unit 1101 can be configured to execute the resource pushing method, the method of training a deep learning model in any other appropriate manner (e.g., by means of firmware).

[0176] The various embodiments of the systems and techniques described above in this specification can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGA), application specific integrated circuits (ASIC), application specific standard products (ASSP), systems on chip (SOC), complex programmable logic devices (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that receives data and instructions from, and transmits data and instructions to, a storage system, at least one input device, and at least one output device.

[0177] The program code for implementing the methods 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, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions / operations specified in the flowchart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package partially on the machine and partially on a remote machine, or entirely on the remote machine or server.

[0178] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0179] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and a pointing device (e.g., a mouse or a trackball) through which the user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, speech input, or tactile input).

[0180] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), and the Internet.

[0181] A computer system can include a client and a server. The client and the server are generally far from each other and typically interact through a communication network. The client - server relationship is created by computer programs running on the respective computers and having a client - server relationship with each other. The server can be a cloud server, or a server of a distributed system, or a server incorporating blockchain.

[0182] It should be understood that various forms of the processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recited in this disclosure can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and no limitation is imposed herein.

[0183] The above - described specific embodiments do not constitute a limitation on the protection scope of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub - combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the protection scope of this disclosure.

Claims

1. A resource pushing method, comprising: Obtaining resource-related features for a target object; Performing intention detection on the target object based on the resource-related features to obtain an extended topic feature, wherein the extended topic represented by the extended topic feature satisfies a semantic difference degree condition with the initial topic for the resource-related features; Determining a target resource from candidate resources based on the extended topic feature and pushing the target resource to the target object.

2. The method according to claim 1, wherein The determining the target resource from candidate resources based on the extended topic feature includes: Performing a demand intention evaluation on the candidate resources based on the extended topic feature and candidate resource features for the candidate resources to obtain an initial demand weight; Updating the initial demand weight based on the similarity between the extended topic feature and the candidate resource features to obtain a demand weight for the candidate resources; Determining the target resource from at least one of the candidate resources based on the demand weight.

3. The method according to claim 2, wherein, The performing a demand intention evaluation on the candidate resources based on the extended topic feature and candidate resource features for the candidate resources includes: Performing a demand intention evaluation on the candidate resources based on interaction scenario features among the extended topic feature, the candidate resource features, and the resource-related features; Wherein the interaction scenario feature characterizes the resource interaction scenario of the target object in a specified period.

4. The method according to claim 1, wherein The performing intention detection on the target object based on the resource-related features to obtain an extended topic feature includes: Performing feature fusion on the resource-related features based on an attention mechanism to obtain an initial topic feature representing the initial topic; Performing intention detection based on the initial topic feature to obtain the extended topic feature; Wherein the target resource is determined from the candidate resources based on the extended topic feature and the initial topic feature.

5. The method according to claim 1, wherein, The resource-related features include at least one of the following: Historical resource features for historical resources, where the historical resources are resources on which the target object has performed interaction operations; Object attribute features related to the target object.

6. A method for training a deep learning model, comprising: Obtaining sample resource-related features for a sample object; Processing the sample resource-related features using a first encoder of the deep learning model to obtain a sample initial topic feature; Processing the sample initial topic feature using a second encoder of the deep learning model to obtain a sample extended topic feature; wherein the sample topic represented by the sample initial topic feature satisfies a semantic similarity condition with the sample initial topic represented by the sample resource-related features; Using the sample extended topic feature and the sample initial topic feature as features of a negative sample pair to train the deep learning model based on a contrast learning mechanism to obtain a trained deep learning model.

7. The method according to claim 6, wherein The deep learning model further includes a second detection network; the training the deep learning model based on the contrast learning mechanism includes: Processing the extended topic feature using the second detection network of the deep learning model to obtain a sample topic feature, Use the extended topic features and the initial topic features of the sample as the features of the negative sample pair, and use the topic features and the initial topic features of the sample as the features of the positive sample pair to train the deep learning model based on the contrast learning mechanism.

8. The method according to claim 7, wherein, Training the deep learning model based on the contrast learning mechanism includes: Determine the first difference information between the extended topic features and the initial topic features of the sample; Based on the first difference information and the second difference information characterizing the difference between the topic features and the initial topic features of the sample, train the deep learning model.

9. A resource pushing device, comprising: A first acquisition module for acquiring resource-related features for a target object; An extended topic feature acquisition module for performing intention detection on the target object based on the resource-related features to obtain extended topic features, where the extended topic characterized by the extended topic features satisfies a semantic difference degree condition with the initial topic for the resource-related features; A first determination module for determining a target resource from candidate resources based on the extended topic features and pushing the target resource to the target object.

10. A device for training a deep learning model, comprising: A second acquisition module for acquiring sample resource-related features for a sample object; A sample initial topic feature acquisition module for processing the sample resource-related features by a first encoder of the deep learning model to obtain sample initial topic features; A sample extended topic feature acquisition module for processing the sample initial topic features by a second encoder of the deep learning model to obtain sample extended topic features; wherein the sample topic characterized by the sample initial topic features satisfies a semantic similarity condition with the sample initial topic characterized by the sample resource-related features; A training module for using the sample extended topic features and the sample initial topic features as the features of the negative sample pair to train the deep learning model based on the contrast learning mechanism to obtain a trained deep learning model.

11. An intelligent agent of artificial intelligence, comprising: An input module for receiving input information; A processing module for determining a target task based on the input information received by the input module, determining a large model based on the target task, and obtaining output information by calling the large model to execute the method according to any one of claims 1 to 5, or by calling the large model to execute the method according to any one of claims 6 to 8; An output module for outputting the output information obtained by the processing module.

12. An electronic device, comprising: At least one processor; And A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable 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 execute the method according to any one of claims 1 to 8.

13. 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 8.

14. 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 8.

Images