CN117034183A

CN117034183A - Data processing method, device, computer equipment, storage medium and program product

Info

Publication number: CN117034183A
Application number: CN202211360629.8A
Authority: CN
Inventors: 苏鑫; 单子非; 陈谦
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-11-02
Filing date: 2022-11-02
Publication date: 2023-11-10

Abstract

The embodiment of the application provides a data processing method, a data processing device, computer equipment, a storage medium and a program product, wherein the method comprises the following steps: object feature data of a target service object is obtained, wherein the object feature data comprises object feature information of the target service object in a target service domain and a reference service domain; carrying out semantic fusion understanding on each object feature information in the object feature data to obtain feature semantic information of the target business object; acquiring content semantic information of media content in a target service domain, and determining the matching degree between a target service object and the media content according to semantic correlation between the feature semantic information and the content semantic information; and carrying out recall processing on the media content in the target service domain according to the matching degree between the target service object and the media content. By adopting the embodiment of the application, the matching accuracy between the business object and the media content can be improved, and the accuracy of the recall result can be further improved.

Description

Data processing method, device, computer equipment, storage medium and program product

Technical Field

The present application relates to the field of computer technology, and in particular, to the field of intelligent recall technology, and more particularly, to a data processing method, a data processing apparatus, a computer device, a computer readable storage medium, and a computer program product.

Background

With the rapid development of computer technology, intelligent recall is widely used in various large business scenarios, such as recommending interesting media content to business objects, mining interesting objects of media content, and the like. The key technical point of intelligent recall is to perform accurate matching between the business object and the media content.

Currently, the matching scheme between the business object and the media content mainly includes the following two types: a first matching scheme, based on the access operation (such as viewing, praise, collection, etc.) of the business object to the historical media content, determining the matching condition between the business object and the new media content; second, the matching between the business object and the media content is determined by matching the interest keywords of the business object with the content keywords of the media content. In the first matching scheme, the access operation of the newly registered business object is missing, or the number of the access operations of the business object with low activity is small, so that the matching accuracy of the first matching scheme to the cold start scene is low, and the accuracy of recall results in the cold start scene is low; the second matching scheme has higher dependency on the keyword construction system, and the subtle gap between the interest keyword construction system of the business object and the content keyword construction system of the media content has larger influence on the matching accuracy. It can be seen that the matching accuracy of the current intelligent recall scheme between the business object and the media content is not high, so that the recall result is not high in accuracy.

Disclosure of Invention

The embodiment of the application provides a data processing method, a data processing device, computer equipment, a storage medium and a program product, which can improve the matching accuracy between a business object and media content and further improve the accuracy of recall results.

In one aspect, an embodiment of the present application provides a data processing method, where the data processing method includes:

acquiring object feature data of a target business object; the object feature data of the target service object comprises object feature information of the target service object in a target service domain and object feature information of the target service object in a reference service domain;

carrying out semantic fusion understanding on each object feature information from different service domains in object feature data of a target service object to obtain feature semantic information of the target service object;

acquiring content semantic information of media content in a target service domain, and determining the matching degree between the target service object and the media content according to semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content;

and carrying out recall processing on the media content in the target service domain according to the matching degree between the target service object and the media content.

Accordingly, an embodiment of the present application provides a data processing apparatus, including:

the acquisition unit is used for acquiring object feature data of the target business object; the object feature data of the target service object comprises object feature information of the target service object in a target service domain and object feature information of the target service object in a reference service domain;

the processing unit is used for carrying out semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object to obtain feature semantic information of the target service object;

the acquisition unit is also used for acquiring the content semantic information of the media content in the target service domain, and determining the matching degree between the target service object and the media content according to the semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content;

and the processing unit is also used for carrying out recall processing on the media content in the target service domain according to the matching degree between the target service object and the media content.

In one implementation manner, the processing unit is configured to perform semantic fusion understanding on each object feature information from different service domains in object feature data of a target service object, and when obtaining feature semantic information of the target service object, is specifically configured to perform the following steps:

Adding global feature flag information in a feature sequence composed of feature information of each object;

vector coding is carried out on the global feature mark information and each object feature information in the object feature information respectively to obtain vector representation of the global feature mark information and vector representation of each object feature information in the object feature information;

and carrying out semantic fusion coding on the vector representation of the global feature flag information and the vector representation of each object feature information in the object feature information to obtain feature semantic information of the target service object.

In one implementation, semantic fusion encoding is performed by a depth semantic encoding module in an object semantic understanding model, the depth semantic encoding module comprises a plurality of semantic encoding layers, and global feature flag information and object feature information correspond to respective semantic encoding units in each semantic encoding layer;

the processing unit is used for carrying out semantic fusion coding on the vector representation of the global feature flag information and the vector representation of each object feature information in the object feature information, and is specifically used for executing the following steps when the feature semantic information of the target service object is obtained:

Calling each semantic coding unit in the first semantic coding layer to respectively code the corresponding information to obtain semantic coding results of each semantic coding unit in the first semantic coding layer;

invoking each semantic coding unit in the second semantic coding layer respectively, carrying out fusion processing on the semantic coding results of each semantic coding unit in the first semantic coding layer, and carrying out semantic coding processing on the fusion processing results of the first semantic coding layer to obtain semantic coding results of each semantic coding unit in the second semantic coding layer;

continuously calling each semantic coding unit in the subsequent semantic coding layer, carrying out fusion processing on the semantic coding result of each semantic coding unit in the previous semantic coding layer of the subsequent semantic coding layer, and carrying out semantic coding processing on the fusion processing result of the previous semantic coding layer to obtain the semantic coding result of each semantic coding unit in the subsequent semantic coding layer;

and carrying out fusion processing on the semantic coding result of each semantic coding unit in the last semantic coding layer to obtain the characteristic semantic information of the target business object.

In one implementation, the obtaining unit is further configured to perform the following steps:

Acquiring access heat information of the media content, wherein the access heat information of the media content is used for reflecting the accessed condition of the media content;

adjusting content semantic information of the media content according to the access heat information of the media content;

the processing unit is used for determining the matching degree between the target service object and the media content according to the semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content, and is specifically used for executing the following steps:

and determining the matching degree between the target business object and the media content according to the semantic correlation between the characteristic semantic information of the target business object and the adjusted content semantic information.

In one implementation, the obtaining unit is configured to, when obtaining content semantic information of media content in the target service domain, specifically perform the following steps:

acquiring content attribute information of media content; the content attribute information comprises content identification of the media content or content description information of the media content, wherein the content identification is used for indicating arrangement positions of the media content in all media content in a target service domain, and the content description information is used for reflecting content characteristics of the media content;

And carrying out semantic understanding on the content attribute information to obtain the content semantic information of the media content in the target service domain.

In one implementation, when there is a content recall requirement for recalling matching media content of a target service object in a target service domain, the processing unit is configured to, according to a degree of matching between the target service object and the media content, perform, when performing recall processing on the media content in the target service domain, the following steps:

if the matching degree between the target service object and the media content meets the matching condition, the media content is used as the matching media content of the target service object;

when there is an object recall requirement for recalling a matched service object of the media content in the target service domain, the processing unit is configured to execute the following steps when performing recall processing on the media content in the target service domain according to the matching degree between the target service object and the media content:

and if the matching degree between the target service object and the media content meets the matching condition, taking the target service object as the matching service object of the media content.

In one implementation, the processing unit is further configured to perform the steps of:

Acquiring characteristic semantic information of a reference service object in a target service domain;

calculating the similarity between the characteristic semantic information of the target business object and the characteristic semantic information of the reference business object;

if the similarity is greater than the similarity threshold, the reference business object is determined to be a similar business object of the target business object.

In one implementation, the semantic fusion encoding is performed by an object semantic understanding model in a recall matching model; content semantic information of the media content is obtained by carrying out semantic understanding on the content attribute information of the media content by a content semantic understanding model in a recall matching model; the processing unit is further used for executing the following steps:

when the object semantic understanding model in the recall matching model has optimization requirements, parameters of the content semantic coding model in the recall matching model are kept unchanged, and the object semantic understanding model is optimized;

when the content semantic coding model in the recall matching model has an optimization requirement, parameters of the object semantic understanding model in the recall matching model are kept unchanged, and the content semantic coding model is optimized.

In one implementation, the data processing method is performed by invoking a recall matching model that includes an object semantic understanding model and a content semantic understanding model; a training process for recalling a matching model, comprising:

Acquiring sample data of a target service domain, wherein the sample data comprises object feature data of a sample service object in the target service domain and a plurality of sample contents; the object feature data of the sample service object comprises object feature information of the sample service object in a target service domain and object feature information of the sample service object in a reference service domain;

invoking an object semantic understanding model in the recall matching model, and carrying out semantic fusion understanding on each object feature information from different service domains in object feature data of the sample service object to obtain feature semantic information of the sample service object;

calling a content semantic understanding model in the recall matching model to respectively acquire content semantic information of each sample content in the plurality of sample contents;

determining the matching degree between the sample service object and the corresponding sample content according to the semantic correlation between the characteristic semantic information of the sample service object and the content semantic information of each sample content;

and determining loss information of the recall matching model according to the matching degree between the sample business object and each sample content, and training the recall matching model according to the loss information.

In one implementation, the processing unit is configured to determine, according to the matching degree between the sample service object and each sample content, loss information of the recall matching model, and specifically configured to perform the following steps:

sequencing each sample content according to the matching degree between the sample business object and each sample content;

determining sample content as a positive sample in the sorting result of each sample content;

and taking the marked sample content of the sample service object as supervision information, and determining the loss information of the recall matching model according to the difference between the marked sample content and the sample content serving as a positive sample.

In one implementation, sample content in the sample data is sampled from a positive sample, which is media content that has been accessed in the target service domain; the processing unit is further used for executing the following steps:

performing access degree reduction processing on a target positive example sample with high access degree in the positive example samples;

or, adjusting the sampling weight of the positive sample accessed by the target access type;

alternatively, the sampling weights of the positive samples belonging to different time intervals are adjusted.

Accordingly, an embodiment of the present application provides a computer apparatus:

A processor adapted to implement a computer program;

a computer readable storage medium storing a computer program adapted to be loaded by a processor and to perform the data processing method described above.

Accordingly, embodiments of the present application provide a computer-readable storage medium storing a computer program which, when read and executed by a processor of a computer device, causes the computer device to execute the above-described data processing method.

Accordingly, embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the data processing method described above.

In the embodiment of the application, in the process of determining the matching degree between the target service object and the media content in the target service domain, the object feature data of the target service object can comprise the object feature information of the target service object in the target service domain and the object feature information of the target service object in the reference service domain, and the situation that the object feature information of the target service object in the target service domain is missing or insufficient in a cold start scene can be compensated by acquiring the object feature information of the target service object in the reference service domain, so that the matching accuracy between the service object and the media content in the cold start scene can be improved; in addition, through carrying out deep semantic understanding on object feature data of the target service object and media content, according to semantic relativity between feature semantic information of the target service object and content semantic information of the media content, the matching degree between the target service object and the media content is determined in semantic dimension, and the matching accuracy between the service object and the media content can be improved; in summary, the embodiment of the application can improve the matching accuracy between the business object and the media content, and further improve the accuracy of the recall result.

Drawings

In order to more clearly illustrate the embodiments of the application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic diagram of Bayesian inference of a data processing scheme provided by an embodiment of the present application;

FIG. 2 is a schematic diagram of a data processing system according to an embodiment of the present application;

fig. 3a is an application scenario schematic diagram of a data processing scheme according to an embodiment of the present application;

fig. 3b is a schematic view of an application scenario of another data processing scheme according to an embodiment of the present application;

FIG. 4 is a schematic flow chart of a data processing method according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a recall matching model according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a semantic coding module according to an embodiment of the present application;

FIG. 7 is a flowchart of another data processing method according to an embodiment of the present application;

FIG. 8 is a schematic diagram of a data processing apparatus according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

The embodiment of the application relates to an intelligent recall technology. The intelligent recall technique is as follows: determining a recall result between the service object and the media content according to the matching condition between the service object and the media content; the intelligent recall technology can be mainly applied to two types of recall scenes, one is a content recall scene, and the other is an object recall scene; wherein, the content recall scene refers to a scene of recalling media content matched with a business object, for example, a scene of recommending the media content of interest to the business object; an object recall scene refers to a scene that recalls a business object that matches media content, e.g., a scene that mines objects of interest of the media content. The media content refers to content carried by media, and in the embodiment of the present application, the media content may include any of text, image, audio, video, and the like in terms of expression, and may include any of advertisement media content, live media content, short video media content, and the like in terms of usage.

Based on the above description of the intelligent recall technique, it is readily apparent that no matter what type of application scenario the intelligent recall technique is applied to, the goal of the intelligent recall technique is to match between a business object and media content to determine the degree of matching between the business object and media content. Aiming at the aim of the intelligent recall technology, a first scheme for matching between a business object and media content based on the access operation of the business object to the media content is provided, and a second scheme for matching between the business object and the media content based on the access operation of the business object to the media content and the content label of the media content is provided; taking an example that the intelligent recall technology is applied to a content recall scene, the following describes a first scheme and a second scheme respectively with reference to fig. 1:

from the perspective of bayesian reasoning, the goal of intelligent recall techniques may be described as P (f|u), where F represents media content and U represents a business object, i.e., estimating whether media content F can be recalled by business object U.

First scheme (corresponding to Step1 in fig. 1):

how to match between the business object and the media content, the probability of recommending a certain media content to the business object next time is judged from the access operation of the existing business object to the media content by Bayesian inference; in a first scenario, the target P (F|U) of the intelligent recall technique is converted to P (U|F) P (F), i.e., P (F|U) ≡ _F∈FΩ P (U|F) P (F) access to media content through existing business objectsThe query operation posterior predicts future access operations. For example, both the service object U1 and the service object U2 like the media content a, the media content B and the media content C at the same time, if the service object U3 already like the media content a and the media content B, then we can recommend the media content C to the service object U3 through the large-scale access operation of the service object. However, in practical use scenarios, the interaction sample of a large number of business objects with media content is very sparse for the whole, especially for business objects of low liveness and new media content; lower sample sampling often leads to larger estimation bias or inefficient generalization of the inference results, based on which a second scheme is proposed.

Second scheme (corresponding to Step2 in fig. 1):

the second scheme introduces a content tag system based on the first scheme, in which the target P (F|U) of intelligent recall technique is converted into P (U|T) P (T|F) P (F), i.e., P (F|U) ≡ _T∈TΩ P (U|T) P (T|F) P (F), wherein T represents a content tag of the media content, the access operation of the media content and the business object to the media content is selected to be corresponding to the content tag, and by means of the popularization capability of the content tag, the media content can be recommended to the matched business object better under the condition of less access operation samples. For example, the low-liveness service object U1 and the low-liveness service object U2 enjoy the media content a and the media content C at the same time, wherein the content label of the media content a is "tennis", and the content label of the media content C is "star XXX"; at this time, if the service object U3 with low liveness accesses the media content B, and the content label of the media content B is also "tennis", we can recommend the media content a and the media content C to the service object U3 according to the relevance between "tennis" and "ball star XXX" with high probability; at this point we no longer need to consider whether media content B is relevant for business object U1 and business object U2, as media content B and media content a have been associated by the content tag "tennis". From this example, it can be seen that access operations by means of content tags can make rational recommendation logic with fewer access operation samples. Although the second scheme The method can overcome the estimation difficulty caused by sample sparsity, but also causes a certain recommendation ambiguity. For example, the content labels of "XXX ball stars seize the crown" and "student tennis team training" can be "tennis", but their actual content is quite different and even ambiguous. Thus, the second scheme is very dependent on the construction of the tag system.

As can be seen from the above description about the first scheme and the second scheme, the first scheme has low accuracy of matching the service object with the media content in the cold start scene, and the second scheme can improve the accuracy of matching the cold start scene to a certain extent, but the second scheme excessively depends on the characteristics of the label system construction, so that the accuracy of matching the service object with the media content in the second scheme is also low.

Based on this, the embodiment of the application provides a data processing scheme, in the recall process of a target service domain, the data processing scheme can acquire the object feature tags of the service object in the target service domain and reference service domain, by performing deep semantic fusion understanding on the object feature tags of the service object in different service domains, the object feature tags of the service object in different service domains can be mapped into the semantic expression space of a high-dimensional vector (emb) and the media content can be mapped into the semantic expression space of the same high-dimensional vector, and then the matching degree between the service object and the media content can be determined according to the semantic correlation between the object feature features of the service object and the content semantics of the media content in the semantic expression space, so that the recall processing of the media content in the target service domain can be performed based on the matching degree between the service object and the media content. As shown in FIG. 1, the data processing scheme (corresponding to Step3 of FIG. 1) according to the embodiment of the present application converts the target P (F|U) of the intelligent recall technique into P (U|T) P (T|E) P (E|F) P (F), i.e., P (F|U)/(C) _T∈TΩ P (U|T) P (T|E) P (E|F) P (F), wherein E represents object feature semantic vectors of business objects in a semantic expression space or content semantic vectors of media content in the semantic expression space, different object features can be linked through the vectors of the semantic expression space, and the object features are minedThe deep association between different service domains makes generalized breakthrough of the label system and the limitation of service range.

According to the data processing scheme provided by the embodiment of the application, on one hand, for a newly registered service object in a target service domain or a service object with medium activity and low activity, the object feature tag of the service object in a reference service domain can be acquired to complement the missing object feature of the service object in the target service domain, so that the matching accuracy between the service object and the media content can be effectively improved in a cold start scene. On the other hand, for the corresponding service object, the high-dimensional semantic expression space can integrate object feature labels associated with the service object, so that the object feature characteristics of the service object are fully expressed, for the media content, the high-dimensional semantic expression space can fully express the content semantics of the media content, and the matching degree between the service object and the media content is determined according to the semantic correlation between the object feature characteristics of the service object and the content semantics of the media content in the semantic expression space, so that the matching accuracy between the service object and the media content can be improved.

The data processing scheme provided by the embodiment of the application relates to the technology of artificial intelligence such as natural language processing, machine learning and the like in the semantic understanding process. Wherein:

artificial intelligence (Artificial Intelligence, AI) is the theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and extend human intelligence, sense the environment, acquire knowledge and use the knowledge to obtain optimal results. In other words, artificial intelligence is an integrated technology of computer science that attempts to understand the essence of intelligence and to produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence, i.e. research on design principles and implementation methods of various intelligent machines, enables the machines to have functions of sensing, reasoning and decision. The artificial intelligence technology is a comprehensive subject, and relates to the technology with wide fields, namely the technology with a hardware level and the technology with a software level. Artificial intelligence infrastructure technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and other directions.

Natural language processing (Nature Language processing, NLP) is an important direction in the fields of computer science and artificial intelligence. It is studying various theories and methods that enable effective communication between a person and a computer in natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Thus, the research in this field will involve natural language, i.e. language that people use daily, so it has a close relationship with the research in linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic questions and answers, knowledge graph techniques, and the like.

Machine Learning (ML) is a multi-domain interdisciplinary, involving multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is specially studied how a computer simulates or implements learning behavior of a human to acquire new knowledge or skills, and reorganizes existing knowledge structures to continuously improve own performance. Machine learning is the core of artificial intelligence, a fundamental approach to letting computers have intelligence, which is applied throughout various areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, confidence networks, reinforcement learning, transfer learning, induction learning, teaching learning, and the like.

A data processing system suitable for implementing the data processing scheme is described below in connection with figure 2.

As shown in fig. 2, the data processing system may include a terminal device 201 and a server 202, and the embodiment of the present application does not limit the number of terminal devices 201, and the number of terminal devices 201 may be one or more; the terminal device 201 and the server 202 may establish a direct communication connection through wired communication or may establish an indirect communication connection through wireless communication, which is not limited in the embodiment of the present application.

The terminal device 201 may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, an intelligent voice interaction device, a smart watch, a vehicle-mounted terminal, an intelligent home appliance, an aircraft, etc.; the server 202 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server for providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network ), basic cloud computing services such as big data and an artificial intelligent platform, which is not limited in the embodiment of the present application;

The application of the data processing scheme provided by the embodiment of the present application in the content recall scene and the object recall scene is described below with reference to the data processing system shown in fig. 2:

content recall scenario:

content recall scenes refer to scenes that recall media content that matches a business object, e.g., recommending media content of interest to a business object, a business object searching for media content of interest, etc. As shown in the content recall scenario schematic diagram in fig. 3a, the terminal device sends a content recall request of the target service object in the target service domain to the server, where the content recall request may be generated and sent when the target service object opens the target service domain in the terminal device, refreshes the target service domain, or searches media content in the target service domain; after receiving the content recall request sent by the terminal equipment, the server selects the matched media content of the target service object from the candidate media contents according to the matching degree between the target service object and each candidate media content in the target service domain, and pushes one or more matched media contents to the terminal equipment.

Object recall scenario:

An object recall scene refers to a scene that recalls a business object that matches media content, e.g., mining and putting an object of interest for advertising media content, mining an object of interest for live video content, and so forth. Taking advertisement media content delivery as an example, as shown in an advertisement delivery scene diagram in fig. 3b, after receiving an advertisement delivery request of an advertiser in a target service domain, a server may respond to the advertisement delivery request, select a matching media object of the advertisement media content from each candidate service object according to a matching degree between the advertisement media content to be delivered and each candidate service object in the target service domain, and deliver the advertisement media content to a terminal device of the matching media object.

Wherein, the matching degree between the business object and the media content can be determined by the following method whether in the content recall scene or the object recall scene: object feature information of the service object in the target service domain and object feature information of the service object in the reference service domain are subjected to semantic fusion understanding to obtain feature semantic information of the service object, and then the matching degree between the service object and the media content can be determined according to semantic correlation between the feature semantic information of the service object and content semantic information of the media content. By carrying out semantic fusion understanding on the object characteristic information of the service object in the reference service domain and the object characteristic information of the service object in the target service domain, the defect that the object characteristic information of the service object is absent in the target service domain or the quantity of the object characteristic information of the service object is small can be overcome when the target service domain is in a cold start scene; in addition, after the object features and the media content are subjected to deep semantics, the matching degree is determined according to the semantic relevance, so that the matching accuracy between the service object and the media content can be improved.

It will be understood that, the data processing system described in the embodiments of the present application is for more clearly describing the technical solution of the embodiments of the present application, and does not constitute a limitation on the technical solution provided by the embodiments of the present application, and those skilled in the art can know that, with the evolution of the system architecture and the appearance of new service scenarios, the technical solution provided by the embodiments of the present application is equally applicable to similar technical problems.

It should be noted that, in the embodiment of the present application, data such as object feature information, object feature labels, etc. related to a service object are related, when the embodiment of the present application is applied to a specific product or technology, permission or consent of the service object needs to be obtained, and collection, use and processing of related data need to comply with related laws and regulations and standards of related countries and regions.

The data processing scheme provided by the embodiment of the application is described in more detail below with reference to fig. 4-7.

The embodiment of the application provides a data processing method, which mainly introduces a semantic understanding process of a service object side, a semantic understanding process of a media content side and a process of determining the matching degree between the service object and the media content. The data processing method may be performed by a computer device, which may be the server 202 in the data processing system shown in fig. 2 described above. As shown in fig. 4, the data processing method may include, but is not limited to, the following steps S401 to S404:

S401, object feature data of a target business object is acquired.

The object feature data of the target service object may include object feature information of the target service object in the target service domain and object feature information of the target service object in the reference service domain. The target business object can be registered in a plurality of business domains, and the target business domain can be any one of the business domains in which the target business object is registered; the reference service domain may be a plurality of service domains in which the target service object is registered, other service domains except the target service domain; alternatively, the reference service domain may be a service domain in which the target service object is registered, other than the target service, the object feature information of which is included in the target service object is more, for example, a service domain in which the number of object feature information included in the target service object exceeds a number threshold, a service domain in which the number of object feature information included in the target service object is ranked by N (N is a positive integer), or the like.

Wherein the target service domain may be a service function module in an application program, so that the target service domain and the reference service domain may be different service function modules in the same application program, for example, the target service domain is a video service function module in a social application program, and the reference service domain may include a game service function module, a news service function module, a live service function module, and the like in the social application program; alternatively, the target service domain may be one application, whereby the target service domain and the reference service domain may be applications of different services, e.g., the target service domain is a video service application, the reference service domain may include a social service application, a game service application, a news service application, and so on.

The object feature information of the target service object refers to information for describing object features of the target service object, and the object feature information may include object interest feature information and/or object attribute feature information; the object interest feature information refers to information for describing object interest features of the target service object, and the object interest feature information may be actively added in each service domain by the target service object, or the object interest feature information may be extracted according to media content of interest of the target service object in each service domain; the object attribute feature information refers to information for describing object attribute features of the target service object, and the object attribute feature information may be actively added by the target service object in each service domain. The existence form of the object feature information can be a feature tag, such as an interest tag or an attribute tag; or, the existence form of the object feature information may also be an entity link relation, where the entity link relation means that the object feature information exists in the form of an entity triplet, and the entity triplet may include two entity words and an association relation between the two entity words, for example, the entity word "football", where the association relation "belongs to" and the entity word "ball" forms an entity triplet.

S402, carrying out semantic fusion understanding on each object feature information from different service domains in object feature data of the target service object to obtain feature semantic information of the target service object.

In the embodiment of the application, the semantic understanding process of the service object side and the semantic understanding process of the media content side are executed by a recall matching model, and the recall matching model is introduced before the semantic understanding process of the service object side and the semantic understanding process of the media content side are introduced: the recall matching model (BERDE) is a two-tower recall model, as shown in fig. 5, in which the left tower of the recall matching model is an object semantic understanding model for executing the semantic understanding process on the business object side, and the right tower of the recall matching model is a content semantic understanding model for executing the semantic understanding process on the media content side.

In step S402, a process of performing semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object, that is, a semantic understanding process on the service object side may be performed by the object semantic understanding model in the recall matching model. As shown in fig. 5, the object semantic understanding model may include: a vector parsing mapping module (BERT Encoder) and a depth semantic coding module (DNN (Deep Neural Networks, depth neural network)); the vector resolution mapping module may be a pre-training language model, for example, BERT (one pre-training language model), transducer (another pre-training language model), and the like, and may be used for vector encoding (or may be referred to as vector resolution mapping) to encode object feature information into a vector representation; the depth semantic coding module may be a depth neural network formed by modifying the pre-training language model, or may be a mainstream depth neural network, for example, LSTM (depth neural network of one mainstream), GPT-3 (depth neural network of another mainstream), and the like, and the depth semantic coding module may be configured to perform semantic fusion coding on a vector representation of object feature information obtained by vector coding by the vector analysis mapping module. The process of performing semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object by the object semantic understanding model to obtain feature semantic information of the target service object specifically comprises the following steps:

The object semantic understanding model is input in the form of a feature sequence composed of the feature information of each object from different service domains, and global feature mark information (for example, the global feature mark information can be a classification mark [ CLS ]), and the global feature mark information can be specifically added in the first position of the feature sequence; the vector analysis mapping module in the object semantic understanding model can be called to respectively carry out vector coding on the global feature mark information and each object feature information in each object feature information to obtain vector representation of the global feature mark information and vector representation of each object feature information in each object feature information; and then, a depth semantic coding module in the object semantic understanding model can be called to carry out semantic fusion coding on the vector representation of the global feature flag information and the vector representation of each object feature information in the object feature information so as to obtain the feature semantic information of the target service object.

In more detail, for the depth semantic coding module, the depth semantic coding module may include a plurality of semantic coding layers, the global feature flag information and the object feature information correspond to respective semantic coding units in each semantic coding layer, and the semantic fusion coding process of the depth semantic coding module may specifically include: calling each semantic coding unit in the first semantic coding layer to respectively code the corresponding information to obtain semantic coding results of each semantic coding unit in the first semantic coding layer; invoking each semantic coding unit in the second semantic coding layer respectively, carrying out fusion processing on the semantic coding results of each semantic coding unit in the first semantic coding layer, and carrying out semantic coding processing on the fusion processing results of the first semantic coding layer to obtain semantic coding results of each semantic coding unit in the second semantic coding layer; continuously calling each semantic coding unit in the subsequent semantic coding layer, carrying out fusion processing on the semantic coding result of each semantic coding unit in the previous semantic coding layer of the subsequent semantic coding layer, and carrying out semantic coding processing on the fusion processing result of the previous semantic coding layer to obtain the semantic coding result of each semantic coding unit in the subsequent semantic coding layer; and carrying out fusion processing on the semantic coding result of each semantic coding unit in the last semantic coding layer to obtain the characteristic semantic information of the target business object.

As shown in fig. 6, after the vector parsing mapping module encodes the global feature flag information and the object feature information into corresponding vector representations (token), the vector representations are input to a depth semantic coding module, the depth semantic coding module shown in fig. 6 includes 3 semantic coding layers, the semantic coding units in the first semantic coding layer are Unit1-X (x=1, 2, …), the semantic coding units in the second semantic coding layer are Unit2-X (x=1, 2, …), the semantic coding units in the third semantic coding layer are Unit3-X (x=1, 2, …), and the semantic coding result of each semantic coding Unit in the third semantic coding layer is fused to obtain the feature semantic information of the target service object. In the embodiment of the application, when the depth semantic coding module can be a depth neural network formed after the pre-training language model is reformed, the first two layers of modules of the pre-training language model (for example, the BERT model) are adopted as the frame basis of the depth semantic coding module, the parameters of the standard pre-training language model are adopted for initialization, the problem of low training reasoning efficiency caused by overlarge whole depth semantic coding module can be relieved by adopting the first two layers of modules, on the one hand, compared with text semantic understanding, the semantic understanding of object feature information needs more meaning of the object feature information, the understanding dimension is relatively thinner, so that the first two layers of modules are more suitable for the scene, the depth semantic coding module can be called BERT-Light, and finally, when global feature mark information is added at the first bit of a feature sequence, the first bit (namely, the corresponding bit of global feature mark information [ CLS ]) of a BERT-Light output vector can be taken to characterize the comprehensive coding (namely, feature semantic information) of each object feature information of a target service object, and the post-setting calculation node of the first bit is subjected to pruning operation, so that the overall understanding efficiency of the whole semantic understanding is accelerated.

It should be added that, the object semantic understanding model is input in the form of a feature sequence composed of the feature information of each object, and global feature flag information [ CLS ] can be added in the feature sequence composed of the feature information of each object, so that after the semantic fusion coding is completed, the coding result corresponding to the global feature flag information [ CLS ] is taken as the feature semantic information of the target service object. In addition, segmentation information (for example, a segmentation marker [ SEP ]) may be inserted into the feature sequence, where the segmentation information may be used to distinguish object feature information from different sources, i.e. from different service domains, on the one hand, and may be used to distinguish different types of object feature information, for example, object attribute feature information and object interest feature information, on the other hand. The semantic fusion understanding process of the feature sequence after the segmentation information is inserted is similar to the semantic fusion understanding process of the feature sequence after the global feature flag information is added, and is not described herein.

In order to improve the matching efficiency between the service object and the media content, when the semantic fusion understanding obtains the feature semantic information of the service object, the feature semantic information of the service object can be stored, and when the feature semantic information of the same service object needs to be obtained, the feature semantic information of the service object can be directly obtained from the storage space without repeatedly obtaining the object feature data to perform the semantic fusion understanding. That is, for the target service object, the feature semantic information of the target service object may be searched in the storage space, if the feature semantic information of the target service object is successfully searched, the feature semantic information of the target service object may be directly obtained from the storage space, if the search fails, steps S401 to S402 may be executed to obtain the object feature data of the target service object, and the feature semantic information of the target service object may be obtained by performing semantic fusion understanding on the object feature information of the target service object from different service domains. By the method, repeated calculation caused by multiple requests of a single service object can be avoided, and the matching efficiency between the service object and the media content is improved.

S403, obtaining content semantic information of the media content in the target service domain, and determining the matching degree between the target service object and the media content according to semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content.

Step S403 may be divided into the following sub-steps S11-S12:

and s11, acquiring content semantic information of the media content in the target service domain.

The media content in the target service domain refers to the media content published in the target service domain, and the content semantic information of the media content can be obtained by semantic understanding of the content attribute information of the media content. Wherein, the content attribute information of the media content can comprise content identification of the media content or content description information of the media content; the content identity may be used to indicate the arrangement position of the media content in all media content in the target service domain, that is, the content identity may be understood as the sequence number of the media content in the target service domain; the content description information may be used to reflect content characteristics of the media content, and may be, for example, any one or more of content keywords of the media content, account information of an issuing account of the media content, title information of the media content, and content category of the media content.

The process of semantic understanding of content attribute information of media content, i.e., the semantic understanding process on the media content side, may be performed by a content semantic understanding model in a recall matching model. The architecture of the content semantic understanding model is related to content specifically contained in the content attribute information of the media content. Specifically:

when content attribute information of the media content includes a content identifier of the media content, a vector parsing mapping module may be included in the content semantic understanding model, and similar to the vector parsing mapping module in the object semantic understanding model, the vector parsing mapping module in the content semantic understanding model may be a pre-training language model, and the vector parsing mapping module may be used for vector encoding (or may be called vector parsing mapping), and encode the content identifier to construct an encoding array, that is, encode the content identifier into a vector representation, and use the vector representation as content semantic information of the media content. That is, the process of semantically understanding content attribute information of media content refers to a process of vector-encoding content identification of media content.

When the content attribute information of the media content comprises the content description information of the media content, the content semantic understanding model is similar to the object semantic understanding model, and the content semantic understanding model can comprise a vector analysis mapping module and a depth semantic coding module; similar to the vector resolution mapping module in the object semantic understanding model, the vector resolution mapping module in the content semantic understanding model may be a pre-trained language model, and the vector resolution mapping module may be used to perform vector encoding (or may be referred to as vector resolution mapping), encode the content description information to construct an encoding array, i.e., encode the content description information into a vector representation; similar to the depth semantic coding module in the object semantic understanding model, the depth semantic coding module in the content semantic understanding model may be a depth neural network formed after the pre-training language model is modified, or may be a mainstream depth neural network, and the depth semantic coding module may be used for performing semantic fusion coding on the vector representation of the content description information obtained by vector coding of the vector analysis mapping module to obtain the content semantic information of the media content. That is, the process of semantic understanding of the content attribute information of the media content refers to the process of semantic fusion understanding of the content description information of the media content, which is similar to the semantic fusion understanding process of the service object side, and specifically, reference may be made to the semantic fusion understanding process of the service object side, which is not described herein.

In order to improve the matching efficiency between the service object and the media content, after the semantic understanding obtains the content semantic information of the media content, the content semantic information of the media content can be stored, and when the content semantic information of the same media content needs to be obtained, the content semantic information can be directly obtained from the storage space without repeatedly obtaining the content attribute information to carry out semantic understanding. That is, the content semantic information of the media content can be searched in the storage space, if the content semantic information of the media content is successfully searched, the content semantic information of the media content can be directly obtained from the storage space, if the content semantic information of the media content is failed to be searched, the content attribute information of the media content can be obtained, and semantic understanding is carried out on the content attribute information of the media content, so that the content semantic information of the media content is obtained. By the method, repeated calculation caused by multiple matching of single media content can be avoided, and matching efficiency between the service object and the media content is improved.

And s12, determining the matching degree between the target service object and the media content according to the semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content.

The feature semantic information of the target service object is a vector in nature, the content semantic information of the media content is also a vector in nature, and the semantic correlation between the feature semantic information of the target service object and the content semantic information of the media content may specifically refer to a distance metric (for example, a Cosine distance metric may be adopted) between the feature semantic information of the target service object and the content semantic information of the media content, and the matching degree between the target service object and the media content may be obtained by performing the distance metric between the feature semantic information of the target service object and the content semantic information of the media content.

It should be added that, in order to improve the matching efficiency between the service object and the media content, besides the content semantic information of the media content, the information such as the heat and quality of the media content can be considered. Based on the above, the access heat information of the media content can be introduced, the access heat information can reflect the accessed condition (such as heat, quality, etc.) of the media content, the content semantic information of the media content can be adjusted according to the access heat information of the media content, and specifically, the access heat information of the media content is multiplied by the content semantic information of the media content. In combination with the foregoing bayesian inference, the content semantic information of the media content may be represented as P (e|f), the access popularity information of the media content may be represented by using the prior probability P (F) of the media content, that is, the content semantic information of the media content after adjustment may be represented as P (e|f) P (F) =p (E, F), the feature semantic information of the target service object may be represented as P (u|t) P (t|e) =p (u|e), and thus, the matching degree between the target service object and the media content may be determined according to the semantic correlation between the feature semantic information of the target service object and the adjusted content semantic information, which may be referred to as the following formula: p (u|t) P (t|e) P (e|f) P (F) =p (u|e) P (E, F).

S404, according to the matching degree between the target service object and the media content, recall processing is carried out on the media content in the target service domain.

In the recall processing process of the media content, different recall scenes have different recall processing processes, and the recall processing processes in the content recall scene and the object recall scene are described below:

in a content recall scenario, that is, when there is a content recall requirement for recalling matching media content of a target service object in a target service domain, according to the matching degree between the target service object and the media content, recall processing is performed on the media content in the target service domain, specifically, it is: if the matching degree between the target service object and the media content meets the matching condition, the media content can be used as the matching media content of the target service object.

In the content recall scenario, determining that the matching degree between the target service object and the media content meets the matching condition may include the following two ways: first, if the matching degree between the target service object and the media content is greater than the matching degree threshold, it may be determined that the matching degree between the target service object and the media content satisfies the matching condition. Second, the media content may be any one of candidate media content to be matched with the target service object in the target service domain, the candidate media content may be ordered according to the order of the matching degree between each candidate media content and the target service object from high to low, and if the media content is arranged before the target position in each candidate media, it may be determined that the matching degree between the target service object and the media content meets the matching condition.

In an object recall scene, namely when an object recall requirement of a matched service object for recalling media content exists in a target service domain, according to the matching degree between the target service object and the media content, the media content is recalled in the target service domain, specifically, the method comprises the following steps of: if the matching degree between the target service object and the media content meets the matching condition, the target service object can be used as the matching service object of the media content.

In the object recall scenario, determining that the matching degree between the target service object and the media content meets the matching condition may include the following two ways: first, if the matching degree between the target service object and the media content is greater than the matching degree threshold, it may be determined that the matching degree between the target service object and the media content satisfies the matching condition. Second, the target service object may be any one of candidate service objects to be matched with the media content in the target service domain, the candidate service objects may be ordered according to the order of the matching degree between each candidate service object and the media content from high to low, and if the target service object is arranged before the reference position in each candidate service object, it may be determined that the matching degree between the target service object and the media content meets the matching condition.

It should be noted that, the feature semantic information of the service object may be extracted and used alone to construct a similarity measure between the service object and the service object, and find similar service objects, so as to implement similar service object diffusion. Specifically, feature semantic information of a reference service object in a target service domain may be acquired, the reference service object is any service object except the target service object in service objects registered in the target service domain, then similarity between the feature semantic information of the target service object and the feature semantic information of the reference service object may be calculated, and if the similarity is greater than a similarity threshold, the reference service object may be used as a similar service object of the target service object. The similar business object diffusion can be applied to various scenes, for example, after the similar business object of the target business object is determined, the matching media content of the similar business object can be used as the matching media content of the target business object, or the media content of the target business object can be used as the matching media content of the similar business object; as another example, after determining similar business objects for the target business object, similar media objects may be recommended to the target business object.

Similarly, content semantic information of the media content can be extracted and used independently to construct a similarity measure between the media content and the media content, and similar media content is found, so that similar media content diffusion can be realized. Specifically, content semantic information of the first media content and content semantic information of the second media content may be acquired, then a similarity between the content semantic information of the first media content and the content semantic information of the second media content may be calculated, and if the similarity is greater than a similarity threshold, it may be determined that the first media content and the second media content are similar media content. The similar media content diffusion can be applied to various scenes, for example, if the first media content is the matching media content of the target service object and the first media content is similar to the second media content, the second media content can be directly used as the matching media content of the target service object, matching between the target service object and the second media content is not needed, and matching efficiency between the service object and the media content is improved.

It should be further noted that, the object feature data of the service object side may change, and new media content may be continuously added in the target service domain, so that the recall matching model needs to be updated and optimized in an embodiment in order to better adapt to the change occurring in the target service domain; in general, the change frequency of the object feature data on the service object side is far lower than that on the media content side, for example, new media content is updated every hour on the media content side, outdated media content is eliminated, the timeliness change is strong, the object feature data of the same service object on the service object side is always longer, the long-term object feature data is probably unchanged for many months, which means that the routine update period requirements of the left tower and the right tower of the recall matching model are greatly different. Based on the above, the embodiment of the application supports the separated local routine updating of the recall matching model, when the object semantic understanding model (i.e. the left tower) in the recall matching model has the optimization requirement, the parameters of the content semantic coding model (i.e. the right tower) in the recall matching model can be kept unchanged, the object semantic understanding model is optimized, and when the content semantic coding model in the recall matching model has the optimization requirement, the parameters of the object semantic understanding model in the recall matching model can be kept unchanged, and the content semantic coding model is optimized. Compared with the global updating of the whole recall matching model, the local instantiation updating of the recall matching model has higher efficiency, and can better adapt to the change of the object feature data and the media content of the service object by updating the recall matching model, thereby improving the matching accuracy between the service object and the media content to a certain extent.

The embodiment of the application provides a data processing method, which mainly introduces a training process of a recall matching model. The data processing method may be performed by a computer device, which may be the server 202 in the data processing system shown in fig. 2 described above. As shown in fig. 7, the data processing method may include, but is not limited to, the following steps S701 to S705:

s701, sample data of a target service domain is acquired.

The sample data of the target business domain may include object feature data and a plurality of sample contents of a sample business object in the target business domain; the object feature data of the sample business object may include object feature information of the sample business object in the target business domain and object feature information of the sample business object in the reference business domain. Similar to the model application phase, the object feature information of the sample business object refers to information describing object features of the sample business object, which may include object interest feature information and/or object attribute feature information.

In the model training stage, sample content in sample data is sampled from a wide variety of positive samples, wherein the positive samples refer to media content which is accessed in a target service domain; in the process of constructing sample data, the distribution of the sample can be adjusted according to the service requirement of the target service domain, so that the matched recall model obtained by training can better meet the service requirement of the target service domain. Wherein, the adjusting the distribution of the positive sample at least can comprise any one or more of the following modes:

The method comprises the steps that access degree reduction processing is carried out on a target positive sample with high access degree in the positive sample, for example, the access degree can be the access frequency, the positive sample with high access degree is accessed 200000 times in a target service domain, the positive sample with low access attention degree is accessed 20 times in the target service domain, so that the frequency of being sampled by the positive sample with high access degree is high in the sampling process of sample content, the head effect is obvious, logarithmic probability downsampling smoothing can be carried out on the positive sample with high access degree, the access degree is reduced, for example, the access times are reduced from 200000 times to 200 times, the head effect can be reduced, and the sampling of sample content is more uniform; alternatively, the sampling weights of the positive examples samples accessed with the target access types may be adjusted, for example, the correlation of certain access types (e.g., praise, favorites, etc.) with the degree of access may be stronger, and the sampling weights of the positive examples samples focused with these access types may be adjusted; or, the sampling weights of the positive samples belonging to different time intervals are adjusted, in general, the positive samples with the closer release time in the target service domain better meet the sampling requirement during model training, and the sampling weights of the positive samples belonging to different time intervals can be adjusted, for example, the closer the time interval is to the current time, the greater the sampling weights of the positive samples belonging to the time interval are, and the farther the time interval is from the current time, the smaller the sampling weights of the positive samples belonging to the time interval are.

S702, calling an object semantic understanding model in the recall matching model, and carrying out semantic fusion understanding on each object feature information from different service domains in object feature data of the sample service object to obtain feature semantic information of the sample service object.

From the foregoing, it can be seen that the recall matching model (bearer) is a two-tower recall model, and as shown in fig. 5, the left tower of the recall matching model is an object semantic understanding model, the object semantic understanding model may be used to execute a semantic understanding process on the service object side, the right tower of the recall matching model is a content semantic understanding model, and the content semantic understanding model may be used to execute a semantic understanding process on the media content side. The recall matching model may support multiple dual-tower build modes, e.g., siamese-bert (sentenceBert) mode of dual-tower sharing parameters, id-casting mode of a cold start scenario, etc. In addition, the training framework of the recall matching model also simultaneously supports different modes such as single machine multi-card, multi-machine multi-card and the like so as to improve the training efficiency of the framework, and the framework has higher universality and application suitability.

In the model training stage, the process of invoking the object semantic understanding model in the recall matching model to perform semantic fusion understanding on the object feature information from different service domains in the object feature data of the sample service object is similar to the process of invoking the object semantic understanding model in the recall matching model to perform semantic fusion understanding on the object feature information from different service domains in the object feature data of the target service object in the model application stage, and the detailed description of the semantic understanding process on the service object side in the model application stage in the embodiment shown in fig. 4 is omitted.

S703, calling a content semantic understanding model in the recall matching model to respectively acquire content semantic information of each sample content in the plurality of sample contents.

In the model training stage, the content semantic information of the sample content may be obtained by performing semantic understanding on the content attribute information of the sample content, and a process of invoking a content semantic understanding model in the recall matching model to perform semantic understanding on the content attribute information of the sample content is similar to a process of invoking a content semantic understanding model in the recall matching model to perform semantic understanding on the content attribute information of the media content in the model application stage, and in particular, reference may be made to the semantic understanding process on the media content side in the embodiment shown in fig. 4 and will not be repeated here.

S704, determining the matching degree between the sample business object and the corresponding sample content according to the semantic correlation between the characteristic semantic information of the sample business object and the content semantic information of each sample content.

The nature of the feature semantic information of the sample service object is a vector, the nature of the content semantic information of the sample content is also a vector, the semantic correlation between the feature semantic information of the sample service object and the content semantic information of the sample content may specifically refer to a distance metric (for example, a Cosine distance metric may be adopted) between the feature semantic information of the sample service object and the content semantic information of the sample content, and the matching degree between the sample service object and the sample content may be obtained by performing the distance metric between the feature semantic information of the sample service object and the content semantic information of the sample content.

In addition, in the model training stage, unlike a common coding normalization mode, the matching recall model adopts a unilateral regularization scheme, namely a left tower for carrying out semantic understanding on a service object side adopts hard regularization of strong normalization, and a right tower for carrying out semantic understanding on media content adopts soft regularization of L2 Loss (a Loss function) constraint training, so that the effect of the following formula is realized: p (u|t) P (t|e) P (e|f) P (F) =p (u|e) P (E, F). As shown in the above formula, the semantic coding of the service object side (i.e. the characteristic semantic information of the service object) under the unilateral operation has strong normalization, so as to express the tendencies of the service object in different directions in the high-dimensional semantic coding space, i.e. the conditional probability P (U|E) of the service object to the semantic coding. And meanwhile, the normalized semantic coding of the business objects can also better expand the application range of the business objects, for example, the similarity diffusion from business object to business object and the like. The right tower for semantic understanding on the media content side needs to consider the heat degree, quality and the like of the right tower besides the matching characteristic of the semantic content, and the information is often expressed in the prior probability P (F) on the media content side. Thus, right tower encoding we fuse the information conditional probability with the prior probability on the media content side, with joint probability P (E, F) as the content encoding of the media content (i.e., content semantic information of the media content). Therefore, the semantic coding output of the right tower does not perform strong normalization, but retains the flexibility that different media contents can take different vector modes, but in order to prevent training from being diverged, we still add the regular training constraint of L2 Loss in the training process, so that the mode value range of the final output vector is controllable.

And S705, determining the loss information of the recall matching model according to the matching degree between the sample business object and each sample content, and training the recall matching model according to the loss information.

Specifically, the sample contents may be sorted according to the matching degree between the sample service object and each sample content, specifically, the sample contents as positive samples may be determined in the sorting result of each sample content according to the order of the matching degree from high to low, and the sample contents as positive samples may be the sample contents arranged in the first place. Then, the marked sample content of the sample service object can be used as supervision information, the loss information of the recall matching model can be determined according to the difference between the marked sample content and the sample content which is used as a positive sample, and then the recall matching model can be trained according to the loss information. In the embodiment of the application, various construction forms of loss information, such as batch softmax (a loss function), global softmax and other construction forms, in which, for sample data of a batch, marked sample contents of a sample business object exist, and in the global softmax construction form, marked sample contents of a sample business object exist.

Training the recall matching model according to the loss information can specifically refer to optimizing model parameters of the recall matching model according to the direction of reducing the loss information so as to train the recall matching model. The "direction in which loss information is reduced" means: model optimization direction with minimum loss information as target; model optimization is performed in the direction, so that loss information generated by the recall matching model after optimization is smaller than loss information generated by the recall matching model before optimization. For example, the loss information of the recall matching model obtained by this calculation is 0.85, and then the loss information generated by the recall matching model after optimization should be less than 0.85.

In addition, in optimizing model parameters of the matching process model. When the content attribute information of the sample content includes the content identifier of the sample content, the content semantic understanding model may include a vector analysis mapping module, and the parameters of the vector analysis mapping module may be updated by adopting a gradient vector descent calculation manner. When the content attribute information of the media content comprises the content description information of the media content, the content semantic understanding model is similar to the object semantic understanding model, the content semantic understanding model can comprise a vector analysis mapping module and a depth semantic coding module, and parameters of the vector analysis mapping module and the depth semantic coding module can be updated and optimized in a gradient back propagation mode.

The foregoing details of the method of embodiments of the present application are provided for the purpose of better implementing the foregoing aspects of embodiments of the present application, and accordingly, the following provides an apparatus of embodiments of the present application.

Referring to fig. 8, fig. 8 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application, where the data processing apparatus may be disposed in a computer device provided in an embodiment of the present application, and the computer device may be the server 202 mentioned in the embodiment of the method. The data processing apparatus shown in fig. 8 may be a computer program (comprising program code) running in a computer device, which may be used to perform some or all of the steps of the method embodiments shown in fig. 4 or fig. 7. Referring to fig. 8, the data processing apparatus may include the following units:

an obtaining unit 801, configured to obtain object feature data of a target service object; the object feature data of the target service object comprises object feature information of the target service object in a target service domain and object feature information of the target service object in a reference service domain;

the processing unit 802 is configured to perform semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object, so as to obtain feature semantic information of the target service object;

The obtaining unit 801 is further configured to obtain content semantic information of media content in the target service domain, and determine a matching degree between the target service object and the media content according to semantic correlation between feature semantic information of the target service object and content semantic information of the media content;

the processing unit 802 is further configured to recall the media content in the target service domain according to the matching degree between the target service object and the media content.

In one implementation manner, the processing unit 802 is configured to perform semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object, and when obtaining feature semantic information of the target service object, specifically is configured to perform the following steps:

the processing unit 802 is configured to perform semantic fusion encoding on the vector representation of the global feature flag information and the vector representation of each object feature information in the respective object feature information, and when obtaining feature semantic information of the target service object, specifically is configured to perform the following steps:

In one implementation, the obtaining unit 801 is further configured to perform the following steps:

the processing unit 802 is configured to determine, according to semantic correlation between feature semantic information of the target service object and content semantic information of the media content, a matching degree between the target service object and the media content, specifically configured to perform the following steps:

In one implementation, the obtaining unit 801 is configured to, when obtaining content semantic information of media content in a target service domain, specifically perform the following steps:

In one implementation, when there is a content recall requirement for recalling matching media content of a target service object in a target service domain, the processing unit 802 is configured to, when performing recall processing on media content in the target service domain according to a matching degree between the target service object and the media content, specifically perform the following steps:

when there is an object recall requirement for recalling a matching service object of the media content in the target service domain, the processing unit 802 is configured to, according to the matching degree between the target service object and the media content, perform the following steps in particular when performing recall processing on the media content in the target service domain:

In one implementation, the processing unit 802 is further configured to perform the following steps:

In one implementation, the semantic fusion encoding is performed by an object semantic understanding model in a recall matching model; content semantic information of the media content is obtained by carrying out semantic understanding on the content attribute information of the media content by a content semantic understanding model in a recall matching model; the processing unit 802 is further configured to perform the following steps:

In one implementation, the processing unit 802 is configured to determine, according to the degree of matching between the sample business object and each sample content, loss information of the recall matching model, and specifically is configured to perform the following steps:

In one implementation, sample content in the sample data is sampled from a positive sample, which is media content that has been accessed in the target service domain; the processing unit 802 is further configured to perform the following steps:

According to another embodiment of the present application, each unit in the data processing apparatus shown in fig. 8 may be separately or completely combined into one or several other units, or some unit(s) thereof may be further split into a plurality of units having smaller functions, which may achieve the same operation without affecting the implementation of the technical effects of the embodiments of the present application. The above units are divided based on logic functions, and in practical applications, the functions of one unit may be implemented by a plurality of units, or the functions of a plurality of units may be implemented by one unit. In other embodiments of the application, the data processing apparatus may also comprise other units, and in practical applications, these functions may also be realized with the assistance of other units, and may be realized by cooperation of a plurality of units.

According to another embodiment of the present application, a data processing apparatus as shown in fig. 7 may be constructed by running a computer program (including program code) capable of executing some or all of the steps involved in the method as shown in fig. 4 or fig. 7 on a general-purpose computing device such as a computer including a processing element such as a Central Processing Unit (CPU), a random access storage medium (RAM), a read only storage medium (ROM), and the like, and a storage element, and implementing the data processing method of the embodiment of the present application. The computer program may be recorded on, for example, a computer-readable storage medium, and loaded into and executed by the computing device described above.

Based on the above-described method and apparatus embodiments, embodiments of the present application provide a computer device, which may be the aforementioned server 202. Referring to fig. 9, fig. 9 is a schematic structural diagram of a computer device according to an embodiment of the application. The computer device shown in fig. 9 includes at least a processor 901, an input interface 902, an output interface 903, and a computer readable storage medium 904. Wherein the processor 901, the input interface 902, the output interface 903, and the computer readable storage medium 904 may be connected by a bus or other means.

The computer readable storage medium 904 may be stored in a memory of a computer device, the computer readable storage medium 904 for storing a computer program comprising computer instructions, and the processor 901 for executing the program instructions stored by the computer readable storage medium 904. The processor 901 (or CPU (Central Processing Unit, central processing unit)) is a computing core and a control core of a computer device, which is adapted to implement one or more computer instructions, in particular to load and execute one or more computer instructions to implement a corresponding method flow or a corresponding function.

The embodiment of the application also provides a computer readable storage medium (Memory), which is a Memory device in the computer device and is used for storing programs and data. It is understood that the computer readable storage medium herein may include both built-in storage media in a computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores an operating system of the computer device. Also stored in the memory space are one or more computer instructions, which may be one or more computer programs (including program code), adapted to be loaded and executed by the processor. Note that the computer readable storage medium can be either a high-speed RAM Memory or a Non-Volatile Memory (Non-Volatile Memory), such as at least one magnetic disk Memory; optionally, at least one computer readable storage medium remotely located from the aforementioned processor.

In some embodiments, one or more computer instructions stored in a computer-readable storage medium 904 may be loaded and executed by the processor 901 to implement the corresponding steps described above with respect to the data processing methods illustrated in fig. 4 or 7. In particular implementations, computer instructions in the computer readable storage medium 904 are loaded by the processor 901 and perform the steps of:

In one implementation, the computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to perform semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object, and when obtaining feature semantic information of the target service object, the method is specifically used for performing the following steps:

the computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to perform semantic fusion encoding on the vector representation of the global feature flag information and the vector representation of each object feature information in the respective object feature information, so as to perform the following steps when obtaining feature semantic information of the target service object:

In one implementation, computer instructions in the computer-readable storage medium 904 are loaded by the processor 901 and are further used to perform the steps of:

computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to determine a degree of matching between the target business object and the media content based on semantic correlation between characteristic semantic information of the target business object and content semantic information of the media content, specifically for performing the steps of:

In one implementation, the computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to obtain content semantic information of media content in a target service domain, specifically for performing the steps of:

In one implementation, when there is a content recall requirement in the target service domain to recall matching media content of the target service object, computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to perform the following steps when the media content is recalled in the target service domain according to the degree of matching between the target service object and the media content:

when there is an object recall requirement for recalling a matching service object of the media content in the target service domain, computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901, and when the recall processing is performed on the media content in the target service domain according to the matching degree between the target service object and the media content, the method specifically is used for executing the following steps:

In one implementation, the semantic fusion encoding is performed by an object semantic understanding model in a recall matching model; content semantic information of the media content is obtained by carrying out semantic understanding on the content attribute information of the media content by a content semantic understanding model in a recall matching model; computer instructions in the computer-readable storage medium 904 are loaded by the processor 901 and further for performing the steps of:

In one implementation, computer instructions in the computer readable storage medium 904 are loaded and executed by the processor 901 to determine loss information for recall matching models based on the degree of matching between sample business objects and respective sample content, and are specifically configured to perform the steps of:

In one implementation, sample content in the sample data is sampled from a positive sample, which is media content that has been accessed in the target service domain; computer instructions in the computer-readable storage medium 904 are loaded by the processor 901 and further for performing the steps of:

According to one aspect of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions so that the computer device performs the data processing methods provided in the various alternatives described above.

The foregoing is merely illustrative of the present application, and the present application is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method of data processing, the method comprising:

Carrying out semantic fusion understanding on each object feature information from different service domains in the object feature data of the target service object to obtain feature semantic information of the target service object;

acquiring content semantic information of media content in the target service domain, and determining the matching degree between the target service object and the media content according to semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content;

2. The method of claim 1, wherein the performing semantic fusion understanding on the object feature information from different service domains in the object feature data of the target service object to obtain feature semantic information of the target service object includes:

adding global feature flag information in a feature sequence composed of the feature information of each object;

vector encoding is carried out on the global feature mark information and each piece of object feature information in the object feature information respectively to obtain vector representation of the global feature mark information and vector representation of each piece of object feature information in the object feature information;

3. The method of claim 2, wherein the semantic fusion encoding is performed by a depth semantic encoding module in an object semantic understanding model, the depth semantic encoding module comprising a plurality of semantic encoding layers, the global feature flag information and the respective object feature information corresponding to respective semantic encoding units in each semantic encoding layer;

the semantic fusion encoding is performed on the vector representation of the global feature flag information and the vector representation of each object feature information in the object feature information to obtain feature semantic information of the target service object, including:

calling each semantic coding unit in a first semantic coding layer to respectively code the corresponding information to obtain semantic coding results of each semantic coding unit in the first semantic coding layer;

invoking each semantic coding unit in a second semantic coding layer respectively, carrying out fusion processing on semantic coding results of each semantic coding unit in the first semantic coding layer, and carrying out semantic coding processing on the fusion processing results of the first semantic coding layer to obtain semantic coding results of each semantic coding unit in the second semantic coding layer;

Continuously calling each semantic coding unit in a subsequent semantic coding layer, performing fusion processing on semantic coding results of all semantic coding units in a previous semantic coding layer of the subsequent semantic coding layer, and performing semantic coding processing on fusion processing results of the previous semantic coding layer to obtain semantic coding results of all semantic coding units in the subsequent semantic coding layer;

4. The method of claim 1, wherein the method further comprises:

the determining the matching degree between the target service object and the media content according to the semantic correlation between the characteristic semantic information of the target service object and the content semantic information of the media content comprises the following steps:

And determining the matching degree between the target service object and the media content according to the semantic correlation between the characteristic semantic information of the target service object and the adjusted content semantic information.

5. The method of claim 1, wherein the obtaining content semantic information of media content in the target service domain comprises:

acquiring content attribute information of the media content; the content attribute information comprises a content identifier of the media content or content description information of the media content, wherein the content identifier is used for indicating arrangement positions of the media content in all media content in the target service domain, and the content description information is used for reflecting content characteristics of the media content;

and carrying out semantic understanding on the content attribute information to obtain content semantic information of the media content in the target service domain.

6. The method of claim 1, wherein when there is a content recall requirement in the target service domain to recall matching media content of the target service object, the recall processing of the media content in the target service domain according to the degree of matching between the target service object and the media content comprises:

when there is an object recall requirement for recalling the matching service object of the media content in the target service domain, the recalling the media content in the target service domain according to the matching degree between the target service object and the media content includes:

7. The method of claim 1, wherein the method further comprises:

acquiring characteristic semantic information of a reference service object in the target service domain;

and if the similarity is greater than a similarity threshold, determining the reference service object as a similar service object of the target service object.

8. The method of claim 1, wherein the semantic fusion understanding is performed by an object semantic understanding model in a recall matching model; the content semantic information of the media content is obtained by semantic understanding of the content attribute information of the media content by a content semantic understanding model in the recall matching model; the method further comprises the steps of:

When the object semantic understanding model in the recall matching model has optimization requirements, maintaining parameters of a content semantic coding model in the recall matching model unchanged, and optimizing the object semantic understanding model;

when the content semantic coding model in the recall matching model has optimization requirements, parameters of an object semantic understanding model in the recall matching model are kept unchanged, and the content semantic coding model is optimized.

9. The method of claim 1, wherein the method is performed by invoking a recall matching model, the recall matching model comprising an object semantic understanding model and a content semantic understanding model; the training process of the recall matching model comprises the following steps:

acquiring sample data of the target service domain, wherein the sample data comprises object feature data of a sample service object in the target service domain and a plurality of sample contents; the object feature data of the sample service object comprises object feature information of the sample service object in the target service domain and object feature information of the sample service object in a reference service domain;

invoking an object semantic understanding model in the recall matching model, and carrying out semantic fusion understanding on each object feature information from different service domains in the object feature data of the sample service object to obtain feature semantic information of the sample service object;

Invoking a content semantic understanding model in the recall matching model to respectively acquire content semantic information of each sample content in the plurality of sample contents;

10. The method of claim 9, wherein the determining the loss information of the recall matching model based on the degree of matching between the sample business object and the respective sample content comprises:

sequencing each sample content according to the matching degree between the sample service object and each sample content;

11. The method of claim 9, wherein sample content in the sample data is sampled from a positive sample, the positive sample being media content that has been accessed in the target service domain; the method further comprises the steps of:

performing access degree reduction processing on a target positive example sample with high access degree in the positive example samples; or,

adjusting the sampling weight of the positive sample accessed by the target access type; or,

the sampling weights of the positive examples belonging to different time intervals are adjusted.

12. A data processing apparatus, the apparatus comprising:

the acquisition unit is used for acquiring object feature data of the target business object; the object feature data of the target service object comprises object feature information of the target service object in the target service domain and object feature information of the target service object in a reference service domain;

The acquiring unit is further configured to acquire content semantic information of media content in the target service domain, and determine a matching degree between the target service object and the media content according to semantic correlation between feature semantic information of the target service object and content semantic information of the media content;

13. A computer device, the computer device comprising:

a processor adapted to implement a computer program;

a computer readable storage medium storing a computer program adapted to be loaded by the processor and to perform the data processing method according to any one of claims 1-11.

14. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program adapted to be loaded by a processor and to perform the data processing method according to any of claims 1-11.

15. A computer program product comprising computer instructions which, when executed by a processor, implement a data processing method as claimed in any one of claims 1 to 11.