Object sorting method and device, model training method and device, medium and equipment

Abstract

The invention provides an object sorting method and device, a model training method and device, a medium and equipment, and relates to the technical field of artificial intelligence, in particular to the technical field of machine learning and information retrieval. According to the implementation scheme, the method comprises the steps of obtaining a candidate list, a current search word corresponding to the candidate list and a historical behavior sequence; based on the historical behavior sequence and the current search word, utilizing a first interaction network in a sorting model to obtain a first fusion feature; for each candidate object in the candidate list, based on the object information of the candidate object and the first fusion feature, predicting to obtain a prediction score of the candidate object; and sorting the candidate objects in the candidate list based on the predicted score of each candidate object in the candidate list.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more particularly to the field of machine learning and information retrieval technology, specifically to an object sorting method based on a sorting model, a training method for the sorting model, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0003] Object retrieval and search result ranking, as core pillars of information retrieval and artificial intelligence, have evolved from a delicate skill into a key technology driving the digital world. Their fundamental task is to efficiently locate relevant information from massive amounts of data, intelligently rank its importance based on complex criteria, and ultimately present the most relevant, authoritative, and user-intended results.

[0004] The methods described in this section are not necessarily methods that had been previously conceived or adopted. Unless otherwise specified, no method described in this section should be assumed to be prior art simply because it is included in this section. Similarly, unless otherwise specified, the issues mentioned in this section should not be considered to be accepted in any prior art. Summary of the Invention

[0005] This disclosure provides an object sorting method based on a sorting model, a training method for the sorting model, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] According to one aspect of this disclosure, an object ranking method based on a ranking model is provided, comprising: obtaining a candidate list, current search terms corresponding to the candidate list, and historical behavior sequences, wherein the historical behavior sequences include historical search terms of a user application and object information of historically clicked objects corresponding to the historical search terms; obtaining a first fusion feature based on the historical behavior sequence and the current search terms, using a first interaction network in the ranking model, wherein the first interaction network is trained based on a generative task for generating object information of real clicked objects by the user; predicting a predicted score for each candidate object in the candidate list based on the object information of the candidate object and the first fusion feature; and ranking the candidate objects in the candidate list based on the predicted scores of each candidate object in the candidate list.

[0007] According to another aspect of this disclosure, a method for training a ranking model is provided. The ranking model includes a first interactive network, comprising: acquiring sample data, the sample data including sample search terms, historical behavior sequences, and object information of user-actually clicked objects corresponding to the sample search terms; acquiring a first feature based on the historical behavior sequences and sample search terms using the first interactive network; processing the first feature using a decoder to generate predicted object information; calculating a generative loss based on the predicted object information and the object information of user-actually clicked objects; and training the first interactive network using the generative loss.

[0008] According to another aspect of this disclosure, an object ranking apparatus based on a ranking model is provided, comprising: a first acquisition unit configured to acquire a candidate list, a current search term corresponding to the candidate list, and a historical behavior sequence, the historical behavior sequence including historical search terms of a user application and object information of historically clicked objects corresponding to the historical search terms; a second acquisition unit configured to acquire a first fusion feature based on the historical behavior sequence and the current search term, using a first interaction network in the ranking model, the first interaction network being trained based on a generative task for generating object information of real clicked objects by the user; a first prediction unit configured to predict a predicted score for each candidate object in the candidate list based on the object information of the candidate object and the first fusion feature; and a ranking unit configured to rank the candidate objects in the candidate list based on the predicted scores of each candidate object in the candidate list.

[0009] According to another aspect of this disclosure, a training apparatus for a ranking model is provided. The ranking model includes a first interactive network, comprising: a first acquisition unit configured to acquire sample data, the sample data including sample search terms, historical behavior sequences, and object information of user-actually-clicked objects corresponding to the sample search terms; a second acquisition unit configured to acquire a first feature based on the historical behavior sequences and sample search terms using the first interactive network; a generation unit configured to process the first feature using a decoder to generate predicted object information; a first calculation unit configured to calculate a generative loss based on the predicted object information and the object information of user-actually-clicked objects; and a first training unit configured to train the first interactive network using the generative loss.

[0010] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure.

[0011] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to execute the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure.

[0012] According to another aspect of this disclosure, a computer program product is provided, including a computer program, wherein the computer program, when executed by a processor, implements the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0015] Figure 1 A schematic diagram of an exemplary system in which the various methods described herein may be implemented according to embodiments of the present disclosure is shown; Figure 2A flowchart is shown of an object sorting method based on a sorting model implemented according to an embodiment of the present disclosure; Figure 3 A schematic diagram of training a first interactive network based on a generative task according to an exemplary embodiment of the present disclosure is shown; Figure 4 A schematic diagram of object similarity cluster mining according to an exemplary embodiment of the present disclosure is shown; Figure 5 A schematic diagram of a process for determining semantic identifiers based on semantic vectors according to an exemplary embodiment of the present disclosure is shown; Figure 6 A schematic diagram illustrating the independent determination of semantic identifiers for each information domain according to an exemplary embodiment of the present disclosure is shown; Figure 7 A schematic diagram illustrating the acquisition of semantic identifier encoding according to an exemplary embodiment of the present disclosure is shown; Figure 8 A schematic diagram of the structure of a second interactive network according to an exemplary embodiment of the present disclosure is shown; Figure 9 A schematic diagram of the architecture of a sorting model according to an exemplary embodiment of the present disclosure is shown; Figure 10 A flowchart illustrating a training method for a ranking model according to an embodiment of the present disclosure is shown; Figure 11 A structural block diagram of an object sorting apparatus based on a sorting model implemented according to an embodiment of the present disclosure is shown; Figure 12 A structural block diagram of a training apparatus for a ranking model according to an embodiment of the present disclosure is shown; Figure 13 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0016] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0017] In this disclosure, unless otherwise stated, the use of terms such as "first," "second," etc., to describe various elements is not intended to limit the positional, temporal, or importance relationships of these elements; such terms are merely used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of that element, while in other cases, based on the context, they may refer to different instances.

[0018] The terminology used in the description of the various examples described in this disclosure is for the purpose of describing particular examples only and is not intended to be limiting. Unless the context explicitly indicates otherwise, an element may be one or more unless the number of elements is specifically limited. Furthermore, the term "and / or" as used in this disclosure covers any one of the listed items and all possible combinations thereof.

[0019] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0020] Figure 1 A schematic diagram of an exemplary system 100 in which the various methods and apparatus described herein can be implemented according to embodiments of this disclosure is shown. Reference Figure 1 The system 100 includes one or more client devices 101, 102, 103, 104, 105 and 106, a server 120, and one or more communication networks 110 coupling the one or more client devices to the server 120. The client devices 101, 102, 103, 104, 105 and 106 can be configured to execute one or more applications.

[0021] In embodiments of this disclosure, server 120 may run one or more services or software applications that enable the execution of the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure.

[0022] In some embodiments, server 120 may also provide other services or software applications, which may include non-virtual and virtual environments. In some embodiments, these services may be provided as web-based services or cloud services, such as to users of client devices 101, 102, 103, 104, 105 and / or 106 under a Software as a Service (SaaS) model.

[0023] exist Figure 1In the configuration shown, server 120 may include one or more components that implement the functions performed by server 120. These components may include software components, hardware components, or combinations thereof that can be executed by one or more processors. Users operating client devices 101, 102, 103, 104, 105, and / or 106 can sequentially interact with server 120 using one or more client applications to utilize the services provided by these components. It should be understood that various different system configurations are possible and may differ from system 100. Therefore, Figure 1 This is an example of a system used to implement the various methods described herein, and is not intended to be limiting.

[0024] Users can use client devices 101, 102, 103, 104, 105, and / or 106 to enter search terms and initiate object searches. The client devices can provide an interface that allows users to interact with the client devices. The client devices can also output information to the user through this interface. Although... Figure 1 Only six client devices are described, but those skilled in the art will understand that this disclosure can support any number of client devices.

[0025] Client devices 101, 102, 103, 104, 105, and / or 106 may include various types of computer devices, such as portable handheld devices, general-purpose computers (such as personal computers and laptops), workstation computers, wearable devices, smart screen devices, self-service terminal devices, service robots, gaming systems, thin clients, various messaging devices, sensors, or other sensing devices. These computer devices can run various types and versions of software applications and operating systems, such as Microsoft Windows, Apple iOS, UNIX-like operating systems, Linux or Linux-like operating systems (such as Google Chrome OS); or include various mobile operating systems, such as Microsoft Windows Mobile OS, iOS, Windows Phone, and Android. Portable handheld devices may include cellular phones, smartphones, tablets, personal digital assistants (PDAs), etc. Wearable devices may include head-mounted displays (such as smart glasses) and other devices. Gaming systems may include various handheld gaming devices, internet-enabled gaming devices, etc. Client devices are capable of executing various applications, such as various internet-related applications, communication applications (such as email applications), short message service (SMS) applications, and can use various communication protocols.

[0026] Network 110 can be any type of network well known to those skilled in the art, and can support data communication using any of a variety of available protocols (including but not limited to TCP / IP, SNA, IPX, etc.). By way of example only, one or more networks 110 can be a local area network (LAN), an Ethernet-based network, a token ring network, a wide area network (WAN), the Internet, a virtual network, a virtual private network (VPN), an intranet, an extranet, a blockchain network, a public switched telephone network (PSTN), an infrared network, a wireless network (e.g., Bluetooth, WIFI), and / or any combination of these and / or other networks.

[0027] Server 120 may include one or more general-purpose computers, special-purpose server computers (e.g., PC (personal computer) servers, UNIX servers, mid-range servers), blade servers, mainframe computers, server clusters, or any other suitable arrangement and / or combination. Server 120 may include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices for servers). In various embodiments, server 120 may run one or more services or software applications that provide the functionality described below.

[0028] The computing unit in server 120 can run one or more operating systems, including any of the aforementioned operating systems and any commercially available server operating system. Server 120 can also run any of a variety of additional server applications and / or middleware applications, including HTTP servers, FTP servers, CGI servers, JAVA servers, database servers, etc.

[0029] In some implementations, server 120 may include one or more applications to analyze and merge data feeds and / or event updates received from users of client devices 101, 102, 103, 104, 105 and / or 106. Server 120 may also include one or more applications to display data feeds and / or real-time events via one or more display devices of client devices 101, 102, 103, 104, 105 and / or 106.

[0030] In some implementations, server 120 can be a server for a distributed system or a server integrated with blockchain. Server 120 can also be a cloud server, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology. A cloud server is a host product in the cloud computing service system, designed to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0031] System 100 may also include one or more databases 130. In some embodiments, these databases may be used to store data and other information. For example, one or more of the databases 130 may be used to store information such as audio files and video files. Databases 130 may reside in various locations. For example, a database used by server 120 may be local to server 120, or it may be located away from server 120 and may communicate with server 120 via a network-based or dedicated connection. Databases 130 may be of different types. In some embodiments, the database used by server 120 may be, for example, a relational database. One or more of these databases may store, update, and retrieve data from and from the databases in response to commands.

[0032] In some embodiments, one or more of the databases 130 may also be used by an application to store application data. The databases used by the application may be of different types, such as key-value stores, object stores, or regular stores supported by a file system.

[0033] Figure 1 The system 100 can be configured and operated in various ways to enable the application of the various methods and apparatus described in this disclosure.

[0034] According to embodiments of this disclosure, such as Figure 2 As shown, an object ranking method based on a ranking model is provided, including: step S201, obtaining a candidate list, the current search term corresponding to the candidate list, and a historical behavior sequence, wherein the historical behavior sequence includes the user's historical search terms and the object information of the historical clicked objects corresponding to the historical search terms; step S202, based on the historical behavior sequence and the current search term, using a first interaction network in the ranking model, obtaining a first fusion feature, wherein the first interaction network is trained based on a generative task used to generate object information of the user's real clicked objects; step S203, for each candidate object in the candidate list, predicting the predicted score of the candidate object based on the object information of the candidate object and the first fusion feature; and step S204, ranking the candidate objects in the candidate list based on the predicted score of each candidate object in the candidate list.

[0035] Therefore, by first using a first interactive network trained on generative tasks to learn the user's real behavioral migration path to process historical behavior sequences and current search terms, it is possible to obtain the first fusion feature representing the user's deep intent. Subsequently, by using this accurate intent representation to perform decoupled matching with candidate objects, the relevance between the two can be evaluated more accurately, significantly improving the overall accuracy of object ranking.

[0036] In some embodiments, the candidate list may be obtained through a recall module. For example, the recall module may, based on the current search term entered by the user, recall a set of candidate objects initially associated with the search term from an object library (e.g., an advertising library) containing a large number of objects to form the candidate list.

[0037] In some embodiments, the historical behavior sequence may include at least one historical search behavior of the user within a certain time window, each historical search behavior including historical search terms previously entered by the user and object information corresponding to each historical clicked object for each historical search term. In some embodiments, the aforementioned historical search behaviors may be organized in chronological order.

[0038] In some embodiments, the object information of historical clicked objects or the object information of candidate objects can be feature data used to characterize an object (such as an advertisement). Taking an advertisement as an example, the object information can include the text information of the advertisement (such as title, category, description text, etc.) and images. The object information can also include feature vectors obtained by processing the above information (e.g., embedding and concatenating).

[0039] In some embodiments, the first interaction network in the ranking model described above can be used to perform deep encoding on the input user historical behavior sequence and current retrieval behavior (i.e., current retrieval term) to capture the user's true intent and behavior migration path in the sequence.

[0040] In some embodiments, the first interaction network may be an encoder based on a Transformer architecture. In some exemplary embodiments, the first interaction network may be a bidirectional attention network.

[0041] In some embodiments, the first interactive network may also be other sequence modeling networks known in the art, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), or gated recurrent unit (GRU) networks. It is understood that those skilled in the art can determine the network structure of the first interactive network according to actual needs, and no restrictions are imposed here.

[0042] In some embodiments, the current search term (i.e. the current user behavior) can be concatenated with each historical behavior (including historical search terms and the object information of each corresponding historical clicked object) to form a user behavior sequence. After embedding, the sequence is input into a first interaction network (such as a Transformer encoder) for bidirectional attention processing to obtain a first fusion feature. The first fusion feature may include feature information corresponding to each user behavior in the user behavior sequence.

[0043] In some embodiments, the first interactive network (such as RNN or LSTM) can first encode the historical behavior sequence to obtain a historical representation; then, the historical representation is fused with the representation of the current search term (e.g., concatenation or attention operation) to obtain a first fused feature.

[0044] In some embodiments, the first interaction network may be obtained by training a generative task based on generating object information for generating real clicked objects by the user.

[0045] Figure 3 A schematic diagram of training a first interactive network based on a generative task according to an exemplary embodiment of the present disclosure is shown.

[0046] During the training phase, the first interactive network acts as an encoder, and its output first fusion feature is used as input to a decoder based on a Transformer architecture. The decoder can generate object information for a predicted object based on the first fusion feature, and perform generative loss calculation with the object information of the user's actual clicked object corresponding to the current search term. The generative loss is then used to adjust the first interactive network and the encoder. Thus, through this generative training task, the first interactive network can deeply capture the semantic co-occurrence relationships and behavioral transfer paths of the current search term, historical search terms in historical behavior, and historical clicked objects. This allows it to accurately extract the first fusion feature that represents the user's current deep intent, improving the accuracy of subsequent object ranking.

[0047] In some embodiments, the object information of the historical clicked object or the object information of the candidate object may be a semantic identifier of the object.

[0048] In some embodiments, object information includes semantic identifiers, and obtaining semantic identifiers may include: generating corresponding semantic vectors using a pre-trained large model based on object description information, wherein the pre-trained large model is obtained by fine-tuning training based on the description information of each object in the object similarity cluster, and the object similarity cluster contains multiple objects corresponding to the same search term; and determining semantic identifiers based on semantic vectors.

[0049] Therefore, by using a specially fine-tuned language model to generate semantic identifiers for objects, and inputting these semantic identifiers as object information for each candidate object and historical clicked object into the ranking model, deep semantic-based matching and reasoning can be achieved within the model, effectively improving the semantic understanding and ranking ability for new objects or long-tail objects.

[0050] In some embodiments, the process of obtaining the semantic identifier is designed to convert the original object description information into a structured, semantically meaningful identifier (ID).

[0051] In some embodiments, object description information may include one or more information fields of the object, such as one or more of the following multimodal information: object title, detailed description, landing page text content, and corresponding image of the object.

[0052] In some embodiments, the acquisition of the semantic identifier relies on a pre-trained large model, which can be a large language model or a multimodal large model.

[0053] To adapt this large model to the business scenario it is applied to (such as recommendation and ad ranking), it needs to be fine-tuned first using "object similarity clusters".

[0054] In some embodiments, object similarity clusters may be obtained by mining common user behaviors. Figure 4 A schematic diagram of object similarity cluster mining according to an exemplary embodiment of the present disclosure is shown. For example... Figure 4 As shown, user co-click behavior may include the same or different users clicking on multiple different objects under the same search term (for example, object 1 "champagne rose" and object 2 "Diana pink rose" are both clicked under the search term "rose"). Thus, it is considered that the above different objects have a semantic similarity relationship, and the above multiple different objects are classified into an object similarity cluster.

[0055] In some embodiments, it is also possible to analyze the same object that is clicked by users under different search terms (for example, the search terms "model 1" and "model 2" both lead to the same advertisement), thereby establishing similarity relationships between search terms, and further expanding object similarity clusters based on the mining of similarity relationships between objects.

[0056] In some embodiments, after obtaining object similarity clusters, the large model can be fine-tuned using contrastive learning. For example, objects in the same similarity cluster can be used as positive sample pairs, and objects in different clusters can be used as negative sample pairs. By using positive and negative sample pairs, the large model can perform contrastive learning, thereby enabling it to generate more similar semantic vectors for object descriptions of different objects in the same similarity cluster, and to generate more different semantic vectors for object descriptions of two objects in different similar clusters.

[0057] After obtaining the finely tuned pre-trained large model, the object description information (such as the object title) can be input into the pre-trained large model to obtain the corresponding semantic vector.

[0058] In some embodiments, the generated semantic vector can be directly determined as the semantic identifier.

[0059] In some embodiments, semantic identifiers may include multi-level identifiers for representing semantic categories from coarse to fine.

[0060] Therefore, by using multi-level identifiers to represent semantic categories from coarse to fine, hierarchical semantic labels are provided for objects, enabling the model to grasp the commonalities of major categories and distinguish the differences of minor categories, thereby achieving more refined semantic management and more accurate granular control, and thus improving the accuracy of object sorting.

[0061] In some embodiments, determining a semantic identifier based on a semantic vector may include: performing multiple rounds of first operations on the semantic vector, wherein each round of first operations includes: determining, among multiple center vectors corresponding to the current round, a target center vector that is closest to the object vector corresponding to the current round; determining the number corresponding to the target center vector as the sub-identifier corresponding to the current round; and the object vector corresponding to the first round as the semantic vector; and determining the residual vector between the object vector corresponding to the current round and the target center vector as the object vector corresponding to the next round, wherein the multiple center vectors corresponding to the current round are obtained by clustering the object vectors corresponding to the current rounds of each object in the object library; and combining the multiple sub-identifiers obtained from the multiple rounds of first operations into a multi-level identifier.

[0062] Figure 5 A schematic diagram of a process for determining semantic identifiers based on semantic vectors according to an exemplary embodiment of the present disclosure is shown.

[0063] In some exemplary embodiments, the semantic identifier of an object title is taken as an example, such as Figure 5 As shown, the semantic identifier of an object title can be set as a three-layer identifier representing semantic categories from coarse to fine. First, the semantic vectors of the object titles of each object in the object library can be obtained through a pre-trained large model. Then, multiple center vectors corresponding to each layer can be determined. First, the semantic vectors of all objects in the object library can be clustered to obtain multiple center vectors 1 in the first layer, and each center vector 1 can be numbered. Next, the first-layer residual vector (residual vector 1) of each object can be calculated, which is the difference between the semantic vector of the object and its nearest center vector 1. Then, the first-layer residual vectors of each object can be clustered again to obtain multiple center vectors, which can be used as multiple center vectors 2 in the second layer, and each center vector 2 can be numbered. Then, based on similar operations, multiple center vectors 3 in the third layer and their respective numbers can be obtained.

[0064] When it is necessary to obtain the semantic identifier of a certain object (which can be an existing object in the object library or a newly added object in the object library), such as Figure 5As shown, the semantic vector of the object can be obtained first, and used as the first-level object vector 1. The target center vector 1 is then retrieved from multiple center vectors 1, and its corresponding number is designated as the first-level sub-identifier (sub-identifier 1). Next, the residual between object vector 1 and target center vector 1, i.e., residual vector 1, is obtained and used as the second-level object vector 2. Based on object vector 2, the target center vector 2 is retrieved from multiple center vectors 2, and its corresponding number is designated as the second-level sub-identifier (sub-identifier 2). Sub-identifier 3 is then obtained through a similar operation. Sub-identifier 1, sub-identifier 2, and sub-identifier 3 together constitute the semantic identifier of the object. This process of obtaining the semantic identifier can be called the semantic vector quantization process.

[0065] Therefore, by generating multi-level semantic labels through multi-round clustering operations of residual vectorization, a hierarchical semantic tree of objects can be automatically and structurally constructed without manual annotation, thus achieving efficient and automated encoding of massive object semantic information.

[0066] In some embodiments, object information may include multiple semantic identifiers corresponding to multiple information fields, and the multiple information fields may include at least two of the following: a search term field, an object title field, an object landing page field, and an object image field.

[0067] Therefore, by generating semantic tags for different information domains (such as titles, landing pages, etc.), the model can comprehensively understand the semantics of objects and search terms from multiple dimensions, forming a three-dimensional semantic representation, thereby improving the accuracy of ranking by utilizing richer cross-information.

[0068] In some embodiments, the number of rounds of the first operation for obtaining semantic identifiers in different information domains and the number of center vectors corresponding to each round of the first operation can be determined according to the semantic granularity of the corresponding information domain.

[0069] Figure 6 A schematic diagram illustrating the independent determination of semantic identifiers for each information domain according to an exemplary embodiment of this disclosure is shown. For example... Figure 6 As shown, for different information domains of an object, the semantic identifiers corresponding to each of the different information domains of an object can be obtained through personalized quantification operations.

[0070] Semantic granularity refers to the fineness of the semantic information carried by an information domain or the complexity of its semantic space. In the application scenarios of this disclosure, the semantic granularity of different information domains may vary significantly. For the object title domain, its semantic granularity is usually finer. For example, in order to attract users to click, the expression of an advertisement title is very diverse and specific, with a large semantic space. In contrast, the semantic granularity of the object landing page domain is usually coarser. For example, an advertisement landing page often carries traffic from multiple different titles, and its content is a general description of multiple specific products or services, which can be regarded as the average or set of the semantics of all its titles.

[0071] In some examples, for object title fields with finer semantic granularity, a relatively large number of center vectors (e.g., K=1000) can be set in each round of quantization to effectively distinguish subtle semantic differences. For object landing page fields with coarser semantic granularity, a relatively small number of center vectors (e.g., K=200) can be set.

[0072] In some examples, for object title fields with finer semantic granularity, more operation rounds (i.e., the number of semantic tagging levels, such as 4 levels) can be set to capture deeper, finer semantics through more residual calculations. For "object landing page fields" with coarser semantic granularity, fewer operation rounds (such as 2 levels) can be set.

[0073] Therefore, by adaptively determining the number of semantic identifier levels and the number of categories at each level based on the semantic granularity characteristics of different information domains, the rational allocation of computing resources and precise control of semantic granularity are achieved. This personalized quantization method avoids the loss of semantic information caused by granularity mismatch when using uniform parameters for quantization (e.g., coarse-grained information is over-splitting or fine-grained information is forcibly merged), thus ensuring that the optimal semantic representation effect can be obtained in different information domains.

[0074] In some embodiments, the current search term and historical search terms can be converted into semantic identifiers of the corresponding search term domains and then input into the ranking model. The semantic identifiers of the search terms can be obtained based on the methods described above, and will not be elaborated upon here.

[0075] Therefore, by converting search terms into semantic identifiers before inputting them into the model, the information processing format within the model is unified, enabling search terms and objects to be efficiently matched and calculated in the same semantic space, simplifying model design and improving computational efficiency.

[0076] In some embodiments, the semantic identifiers of each search term and each object can be directly input into the ranking model. In some embodiments, the semantic identifiers can also be embedded before being input into the ranking model.

[0077] In some embodiments, each semantic identifier may be processed as follows before being input into the ranking model: embedding each identifier in the multi-level identifier of each semantic identifier to obtain the identifier code corresponding to each identifier; and inputting at least the multiple identifier codes corresponding to the multi-level identifier into a merger to obtain a compressed semantic identifier code, which is then input into the ranking model.

[0078] In some embodiments, in order to address the problem that the input sequence is too long due to the semantic identifier composed of multi-level identifiers, the semantic identifier can be encoded and compressed first, and then input into the sorting model (e.g., the first interactive network).

[0079] In some embodiments, the merger can be a pooling layer. For example, the multiple identifiers can be encoded element-wise by average pooling, or they can be element-wise summed by summation pooling, thereby compressing them into a single vector.

[0080] Therefore, by embedding and compressing multi-level identifiers, high-dimensional sparse identifier sequences can be transformed into low-dimensional dense semantic vectors, reducing the computational complexity of the model while retaining key hierarchical semantic information, which is convenient for subsequent neural network processing.

[0081] In some embodiments, the above-mentioned merger can be constructed based on a Transformer network. At least inputting multiple identifier codes corresponding to multi-level identifiers into the merger to obtain compressed semantic identifier codes may include: determining the position codes corresponding to each identifier code in the multi-level identifier codes according to the hierarchical order of the identifiers in the multi-level identifiers; and inputting multiple identifier codes and the position codes corresponding to each identifier code into the merger to obtain semantic identifier codes.

[0082] Figure 7 A schematic diagram illustrating the acquisition of semantic identifier encoding according to an exemplary embodiment of the present disclosure is shown.

[0083] In some exemplary embodiments, such as Figure 7 As shown, each identifier in the semantic identifier can be embedded to obtain the corresponding identifier codes C1, C2, and C3. Based on the hierarchical order of the identifiers, the positional codes for each identifier code C1, C2, and C3 are determined to be 0, 1, and 2, respectively. Then, each identifier code and its corresponding score positional code can be input into a merger built on a Transformer network. This merger uses a self-attention mechanism to learn the complex interaction relationships between codes at different levels, generating more information-rich compressed codes.

[0084] Therefore, by embedding and compressing multi-level identifiers, high-dimensional sparse identifier sequences (e.g., 3 identifier codes) are transformed into low-dimensional dense single semantic vectors (1 identifier code), which significantly reduces the computational complexity of subsequent ranking models. At the same time, the key hierarchical semantic information is preserved through the merger, which facilitates subsequent neural network processing.

[0085] In some embodiments, the ranking model may further include a second interaction network and a prediction network. Predicting the predicted score of a candidate object based on the object information and the first fusion feature of the candidate object may include: obtaining a second fusion feature of the candidate object using the second interaction network based on the object information and the first fusion feature of the candidate object; and obtaining the predicted score of the candidate object using the prediction network based on the second fusion feature.

[0086] Therefore, by introducing a second interaction network to allow for secondary interaction between candidate information and the first fused features, and then outputting a score via a prediction network, this hierarchical processing mechanism can more precisely characterize the complex relationship between user intent and candidate objects, thereby further improving the accuracy and reliability of the ranking results.

[0087] In some embodiments, the second interaction network can be used to perform deep feature interaction and fusion on a first fusion feature representing the user's deep intent and object information representing the current candidate object.

[0088] In some embodiments, the second interaction network may be, for example, a vector concatenation module that concatenates the encoded vector of the candidate object's information with the first fusion feature in a dimension to form a wider vector as the second fusion feature; the second interaction network may also be, for example, a multilayer perceptron that first concatenates or element-wise multiplies the two input vectors, and then performs nonlinear transformations through several fully connected layers to finally output the second fusion feature.

[0089] Subsequently, based on the second fusion feature, the prediction network can be used to obtain the predicted score of the candidate object. The prediction network can be used to map the high-dimensional fusion feature (second fusion feature) output by the second interaction network to a scalar predicted value. In one embodiment, the prediction network can be a multilayer perceptron, whose final output layer (e.g., a neuron with a sigmoid activation function) is used to output the predicted score.

[0090] In some embodiments, the predicted score is a scalar value used to quantify the probability that the candidate object will be interacted with by the user. For example, the predicted score could be the predicted click-through rate (CTR) or the predicted conversion rate (CVR) of the candidate object.

[0091] In some embodiments, the object sorting method based on the sorting model described above may further include: obtaining the user's context information; and wherein, obtaining the second fusion feature of the candidate object using a second interaction network based on the candidate object's object information and the first fusion feature may include: obtaining the second fusion feature of the candidate object using a second interaction network based on the candidate object's object information, context information, and the first fusion feature.

[0092] In some embodiments, a user's contextual information may include, for example, the user's age, gender, interests, etc. It should be noted that the aforementioned contextual information was obtained after clearly informing the user and obtaining their explicit authorization, and all such information complies with relevant laws and regulations and does not violate public order and good morals.

[0093] Therefore, by further incorporating the user's contextual information into the feature interaction fusion process of the second interaction network, the ranking decision not only depends on the user's historical behavior and current intent, but also comprehensively considers the user's relevant preference information, thereby generating a more personalized ranking result that better meets the user's needs.

[0094] In some embodiments, the second interaction network may include a first interaction subnetwork and a second interaction subnetwork. Based on the object information, context information and first fusion features of the candidate object, obtaining the second fusion feature of the candidate object using the second interaction network may include: using the first interaction subnetwork to perform feature interaction on the object information and context information of the candidate object to obtain a third fusion feature; and using the second interaction subnetwork to obtain the second fusion feature based on the third fusion feature and the first fusion feature.

[0095] The first interaction subnetwork is used to first fuse the object information and context information of the candidate object to generate a personalized candidate object representation. In some embodiments, the first interaction subnetwork can be a deep cross network (DCN), which is configured to explicitly and layer-by-layer compute high-order feature interactions between the candidate object information and the context information. In some embodiments, the first interaction subnetwork can also be a product-based neural network (PNN) based on dot products or Hadamard products, or a standard multilayer perceptron, without limitation.

[0096] The second interaction subnetwork is used to obtain a second fusion feature that reflects the correlation between the personalized candidate object representation (third fusion feature) generated in the previous step and the first fusion feature representing the user's deep intent. In some embodiments, the second interaction subnetwork can be a structure for fusing two vectors, such as a multilayer perceptron or a bilinear interaction network.

[0097] Thus, by splitting the second interaction network into a first interaction sub-network that processes object and context information, and a second interaction sub-network that processes its output and the first fused features, hierarchical and orderly fusion of information is achieved, enabling the model to more effectively integrate multi-source information and improve the quality of feature representation.

[0098] In some embodiments, the second interaction subnetwork may include multiple adaptation subnetworks and a third interaction subnetwork. Obtaining a second fusion feature using the second interaction subnetwork based on a third fusion feature and a first fusion feature may include: inputting the third fusion feature into the multiple adaptation subnetworks respectively to obtain multiple fourth fusion features accordingly, wherein the multiple adaptation subnetworks may be configured to generate multiple fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and for each of the multiple fourth fusion features, inputting the fourth fusion feature and the first fusion feature into the third interaction subnetwork to obtain a fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature may include the fifth fusion feature corresponding to each of the multiple fourth fusion features.

[0099] Therefore, by using multiple adaptor subnetworks to analyze the third fusion feature from different dimensions and interact with the first fusion feature respectively, the model can evaluate the degree of matching between the candidate object and the user's intent from multiple perspectives, thereby capturing complex user interests more comprehensively and meticulously and enhancing the accuracy of ranking.

[0100] In some embodiments, the third interaction subnetwork can be a cross-attention network. Therefore, by applying a cross-attention network as the third interaction subnetwork, the model can dynamically and selectively focus on the parts of the first fusion feature most relevant to the current candidate when generating the second fusion feature, achieving more intelligent and efficient information extraction and fusion.

[0101] Figure 8 A schematic diagram of the structure of a second interactive network according to an exemplary embodiment of the present disclosure is shown.

[0102] In some exemplary embodiments, to achieve more detailed and multi-faceted matching evaluation, the second interaction sub-network is further designed to include multiple adaptation sub-networks (such as...). Figure 8 The network consists of three subnetworks: adapter subnetwork 1, adapter subnetwork 2, adapter subnetwork 3, and a third interactive subnetwork.

[0103] like Figure 8As shown, the third fusion feature (i.e., the "personalized candidate representation") obtained in the previous step can first be input into multiple adaptation sub-networks in parallel. Each adaptation sub-network is configured to focus the candidate on different dimensions or aspects of the first fusion feature (user intent representation). For example, fourth fusion feature 1 can be trained to primarily focus on signals related to the user's recent behavior in the first fusion feature; while fourth fusion feature 2 (generated by adaptation sub-network 2) can be trained to primarily focus on signals related to the user's current search behavior (i.e., the current search term) in the first fusion feature.

[0104] Subsequently, for each fourth fusion feature, it is used as a query vector. The key and value vectors obtained by linearly transforming this fourth fusion feature are then input into the third interaction sub-network to perform cross-attention calculation, thereby obtaining the fifth fusion feature corresponding to each fourth fusion feature. For example, when the input is fourth fusion feature 1 (a query vector focusing on recent behavior) and the key and value vectors obtained by transforming the first fusion feature, it will calculate and output the fifth fusion feature obtained by the interaction and fusion of personalized candidate object representations and representations related to the user's recent behavior.

[0105] The second fusion feature mentioned above is composed of multiple fifth fusion features that represent matching results from different perspectives.

[0106] Therefore, by using multiple adaptive sub-networks to generate multiple query vectors (fourth fusion features) that focus on different aspects of user intent (first fusion feature) (such as recent behavior and current retrieval behavior), and by utilizing a third interactive sub-network (e.g., cross-attention) to perform parallel attention computation, the model can evaluate the relevance of candidate objects to different components of user intent from multiple perspectives. This multi-angle evaluation mechanism can capture complex user interests more comprehensively and meticulously, significantly enhancing the accuracy of the final ranking. The application of cross-attention networks enables each query (fourth fusion feature) from a specific perspective to dynamically and selectively focus on the most relevant parts of user intent (first fusion feature), achieving more intelligent and efficient information extraction and fusion.

[0107] In some embodiments, the ranking model may further include an adaptation network, and the object ranking method based on the ranking model may further include: inputting a first fusion feature into the adaptation network to obtain an adapted first fusion feature; and wherein, based on the second fusion feature, using a prediction network to obtain the prediction score of the candidate object may include: based on the second fusion feature and the adapted first fusion feature, using a prediction network to obtain the prediction score of the candidate object.

[0108] Therefore, by using an adaptation network to adapt the first fusion feature before combining it with the second fusion feature before predicting the score, the contribution of user intent features in the final decision is optimized, which helps to improve the accuracy of the predicted score and makes the final ranking closer to the user's true preferences.

[0109] Figure 9 A schematic diagram of the architecture of a sorting model according to an exemplary embodiment of the present disclosure is shown.

[0110] In some exemplary embodiments, to further improve the accuracy of the final predicted score, the ranking model may also include at least one adaptation network. For example... Figure 9 As shown, the adaptation network (i.e., adaptation network 2) can be configured on the path from the first fused feature to the prediction network.

[0111] In some exemplary embodiments, such as Figure 9 As shown, an adaptation network 1 can be used to adapt the first fused feature before it is input into the third interaction sub-network. The adapted first fused feature is then input into the third interaction sub-network for subsequent feature interaction and fusion. By introducing adaptation network 1, the original user intent representation (first fused feature) can be specifically and optimally adapted to its role (i.e., Key and Value) in cross-attention calculation. This helps improve the accuracy of attention mechanism calculation, enabling candidates to more effectively retrieve the most relevant information from user intent, thereby improving the quality of the final matching features (i.e., each fifth fused feature).

[0112] In some embodiments, the various adaptation subnetworks or adaptation networks described above may be implemented based on fully connected networks or multilayer sensing networks, without limitation. Figure 9 The network structure of the other parts of the ranking model shown has been explained above and will not be repeated here.

[0113] In some embodiments, such as Figure 10 As shown, a training method for a ranking model is provided. The ranking model includes a first interactive network, comprising: step S1001, acquiring sample data, which includes sample search terms, historical behavior sequences, and object information of the user's actual clicked objects corresponding to the sample search terms; step S1002, acquiring a first feature based on the historical behavior sequences and sample search terms using the first interactive network; step S1003, processing the first feature using a decoder to generate predicted object information; step S1004, calculating a generative loss based on the predicted object information and the object information of the user's actual clicked objects; and step S1005, training the first interactive network using the generative loss.

[0114] Therefore, by using generative tasks (such as predicting clicked object information) to train the first interaction network, the network learns the user's true intent and behavior transfer path from historical behavior, laying a solid foundation for the accuracy of subsequent ranking tasks.

[0115] The training method for the first interactive network has been described in detail above and will not be repeated here.

[0116] In some embodiments, the ranking model may further include a second interaction network and a prediction network, and the sample data may further include candidate objects corresponding to sample search terms and labels indicating whether the candidate objects have been clicked. The training method of the ranking model may further include: determining that the first interaction network has completed pre-training in response to the determination that the first interaction network has reached a first convergence condition based on generative loss training; fixing the parameters of the pre-trained first interaction network; obtaining a first fusion feature based on historical behavior sequences and sample search terms using the pre-trained first interaction network; obtaining a second fusion feature of the candidate object based on the object information of the candidate object and the first fusion feature using the second interaction network; obtaining a prediction score of the candidate object based on the second fusion feature using the prediction network; and training the ranking model based on the ranking loss calculated based on the prediction score and the label.

[0117] Therefore, the ranking model is trained by first pre-training the first interaction network and then fixing its parameters. This two-stage training strategy ensures the stability of user intent features, prevents them from being skewed in the early stages of ranking training, and makes the model training process more stable and efficient.

[0118] In some embodiments, the sample data may further include user context information, and wherein obtaining the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information and the first fusion feature may include: obtaining the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information, context information, and the first fusion feature.

[0119] In some embodiments, the second interaction network may include a first interaction subnetwork and a second interaction subnetwork. Based on the object information, context information and first fusion features of the candidate object, obtaining the second fusion feature of the candidate object using the second interaction network may include: using the first interaction subnetwork to perform feature interaction on the object information and context information of the candidate object to obtain a third fusion feature; and using the second interaction subnetwork to obtain the second fusion feature based on the third fusion feature and the first fusion feature.

[0120] In some embodiments, the second interaction subnetwork may include multiple adaptation subnetworks and a third interaction subnetwork. Obtaining a second fusion feature using the second interaction subnetwork based on a third fusion feature and a first fusion feature may include: inputting the third fusion feature into the multiple adaptation subnetworks respectively to obtain multiple fourth fusion features accordingly, wherein the multiple adaptation subnetworks are configured to generate multiple fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and for each of the multiple fourth fusion features, inputting the fourth fusion feature and the first fusion feature into the third interaction subnetwork to obtain a fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature includes the fifth fusion feature corresponding to each of the multiple fourth fusion features.

[0121] In some embodiments, the ranking model may further include an adaptation network, and the training method of the ranking model may further include: inputting a first fusion feature into the adaptation network to obtain an adapted first fusion feature; and wherein, based on the second fusion feature, using a prediction network to obtain the predicted score of the candidate object may include: based on the second fusion feature and the adapted first fusion feature, using a prediction network to obtain the predicted score of the candidate object.

[0122] The network structure and functions of each part of the ranking model described above are similar to those of the ranking model described above, and will not be repeated here.

[0123] In some embodiments, the training method of the ranking model may further include: in response to determining that the ranking model has completed preliminary training by judging that the ranking model has reached a second convergence condition based on the ranking loss training; unfixing the parameters of the pre-trained first interactive network; and jointly fine-tuning each network of the ranking model based on sample data, wherein the first interactive network adopts a first learning rate, and each network in the ranking model other than the first interactive network adopts a second learning rate, and the first learning rate is lower than the second learning rate.

[0124] Therefore, after training the ranking model with the parameters of the first interaction network fixed, the parameters of the first interaction network are further de-fixed, and the various networks in the ranking model are jointly fine-tuned. A lower learning rate is set for the first interaction network, which enables the entire model to be further optimized and adapted, improving overall performance while avoiding forgetting the ability to extract the user's true intent.

[0125] In the training method of the above ranking model, the object information input to the model can be any of the object information mentioned above, such as semantic identifiers, which will not be elaborated here.

[0126] In some embodiments, such as Figure 11As shown, an object ranking device 1100 based on a ranking model is provided, comprising: a first acquisition unit 1110 configured to acquire a candidate list, the current search term corresponding to the candidate list, and a historical behavior sequence, wherein the historical behavior sequence includes the historical search terms of the user application and the object information of the historical clicked objects corresponding to the historical search terms; a second acquisition unit 1120 configured to acquire a first fusion feature based on the historical behavior sequence and the current search term, using a first interaction network in the ranking model, wherein the first interaction network is trained based on a generative task for generating object information of the user's real clicked objects; a first prediction unit 1130 configured to predict a predicted score for each candidate object in the candidate list based on the object information of the candidate object and the first fusion feature; and a ranking unit 1140 configured to rank the candidate objects in the candidate list based on the predicted scores of each candidate object in the candidate list.

[0127] The operations performed by units 1110 to 1140 in the object sorting device 1100 based on the sorting model and the effects they can achieve are similar to steps S201 to S204 in the object sorting method based on the sorting model, and will not be described in detail here.

[0128] In some embodiments, the ranking model may further include a second interaction network and a prediction network. The first prediction unit may include: a first acquisition subunit configured to acquire a second fusion feature of the candidate object using the second interaction network based on the object information and the first fusion feature of the candidate object; and a second acquisition subunit configured to acquire a prediction score of the candidate object using the prediction network based on the second fusion feature.

[0129] In some embodiments, the object sorting apparatus based on the sorting model may further include: a third acquisition unit configured to acquire the user's context information; and wherein the first acquisition subunit may be further configured to: acquire the second fusion feature of the candidate object using a second interaction network based on the candidate object's object information, context information, and first fusion feature.

[0130] In some embodiments, the second interaction network may include a first interaction subnetwork and a second interaction subnetwork, and the first acquisition subunit may be further configured to: use the first interaction subnetwork to perform feature interaction on the object information and context information of the candidate object to obtain a third fusion feature; and use the second interaction subnetwork to obtain a second fusion feature based on the third fusion feature and the first fusion feature.

[0131] In some embodiments, the second interaction subnetwork may include multiple adaptation subnetworks and a third interaction subnetwork. Obtaining a second fusion feature using the second interaction subnetwork based on a third fusion feature and a first fusion feature may include: inputting the third fusion feature into the multiple adaptation subnetworks respectively to obtain multiple fourth fusion features accordingly, wherein the multiple adaptation subnetworks are configured to generate multiple fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and for each of the multiple fourth fusion features, inputting the fourth fusion feature and the first fusion feature into the third interaction subnetwork to obtain a fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature includes the fifth fusion feature corresponding to each of the multiple fourth fusion features.

[0132] In some embodiments, the third interactive subnetwork can be configured as a cross-attention network.

[0133] In some embodiments, the ranking model may further include an adaptation network, and the object ranking apparatus based on the ranking model may further include: a fourth acquisition unit configured to input the first fusion feature into the adaptation network to obtain the adapted first fusion feature; and wherein the second acquisition subunit may be further configured to: obtain the prediction score of the candidate object using a prediction network based on the second fusion feature and the adapted first fusion feature.

[0134] In some embodiments, object information may include semantic identifiers, and the acquisition of semantic identifiers includes: generating corresponding semantic vectors based on object description information using a pre-trained large model, wherein the pre-trained large model is obtained by fine-tuning training based on the description information of each object in the object similarity cluster, and the object similarity cluster contains multiple objects corresponding to the same search term; and determining semantic identifiers based on semantic vectors.

[0135] In some embodiments, semantic identifiers include multi-level identifiers for representing semantic categories from coarse to fine.

[0136] In some embodiments, determining a semantic identifier based on a semantic vector may include: performing multiple rounds of first operations on the semantic vector, wherein each round of first operations includes: determining, among multiple center vectors corresponding to the current round, a target center vector that is closest to the object vector corresponding to the current round; determining the number corresponding to the target center vector as the sub-identifier corresponding to the current round; and the object vector corresponding to the first round as the semantic vector; and determining the residual vector between the object vector corresponding to the current round and the target center vector as the object vector corresponding to the next round, wherein the multiple center vectors corresponding to the current round are obtained by clustering the object vectors corresponding to the current rounds of each object in the object library; and combining the multiple sub-identifiers obtained from the multiple rounds of first operations into a multi-level identifier.

[0137] In some embodiments, object information may include multiple semantic identifiers corresponding to multiple information fields, and the multiple information fields may include at least two of the following: a search term field, an object title field, an object landing page field, and an object image field.

[0138] In some embodiments, the number of rounds of the first operation for obtaining semantic identifiers in different information domains and the number of center vectors corresponding to each round of the first operation can be determined according to the semantic granularity of the corresponding information domain.

[0139] In some embodiments, the current search term and historical search terms can be converted into semantic identifiers of the corresponding search term domains and then input into the ranking model.

[0140] In some embodiments, each semantic identifier may be processed as follows before being input into the ranking model: embedding each identifier in the multi-level identifier of each semantic identifier to obtain the identifier code corresponding to each identifier; and inputting at least the multiple identifier codes corresponding to the multi-level identifier into a merger to obtain a compressed semantic identifier code, which is then input into the ranking model.

[0141] In some embodiments, the above-mentioned merger can be constructed based on a Transformer network. At least inputting multiple identifier codes corresponding to multi-level identifiers into the merger to obtain compressed semantic identifier codes may include: determining the position codes corresponding to each identifier code in the multi-level identifier codes according to the hierarchical order of the identifiers in the multi-level identifiers; and inputting multiple identifier codes and the position codes corresponding to each identifier code into the merger to obtain semantic identifier codes.

[0142] In some embodiments, such as Figure 12 As shown, a training device 1200 for a ranking model is provided. The ranking model includes a first interactive network. The device 1200 includes: a first acquisition unit 1210 configured to acquire sample data, including sample search terms, historical behavior sequences, and object information of user-actually clicked objects corresponding to the sample search terms; a second acquisition unit 1220 configured to acquire a first feature based on the historical behavior sequences and sample search terms using the first interactive network; a generation unit 1230 configured to process the first feature using a decoder to generate predicted object information; a first calculation unit 1240 configured to calculate a generative loss based on the predicted object information and the object information of user-actually clicked objects; and a first training unit 1250 configured to train the first interactive network using the generative loss.

[0143] The operations performed by units 1210 to 1250 in the training device 1200 of the above-mentioned sorting model and the effects they can achieve are similar to steps S1001 to S1005 in the training method of the above-mentioned sorting model, and will not be described in detail here.

[0144] In some embodiments, the ranking model may further include a second interaction network and a prediction network, and the sample data may further include candidate objects corresponding to sample search terms and labels indicating whether the candidate objects have been clicked. The training device for the ranking model may further include: a first determining unit configured to determine that the first interaction network has completed pre-training in response to determining that the first interaction network has reached a first convergence condition based on generative loss training; a fixing unit configured to fix the parameters of the pre-trained first interaction network; a third obtaining unit configured to obtain a first fusion feature based on historical behavior sequences and sample search terms using the pre-trained first interaction network; a fourth obtaining unit configured to obtain a second fusion feature of the candidate object based on the object information of the candidate object and the first fusion feature using the second interaction network; a fifth obtaining unit configured to obtain a prediction score of the candidate object based on the second fusion feature using the prediction network; and a second training unit configured to train the ranking model based on the ranking loss calculated by the prediction score and the label.

[0145] In some embodiments, the sample data may further include user context information, and the fourth acquisition unit may be further configured to: acquire the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information, context information, and first fusion feature.

[0146] In some embodiments, the second interaction network may include a first interaction subnetwork and a second interaction subnetwork, and the fourth acquisition unit may be further configured to: use the first interaction subnetwork to perform feature interaction on the object information and context information of the candidate object to obtain a third fusion feature; and use the second interaction subnetwork to obtain a second fusion feature based on the third fusion feature and the first fusion feature.

[0147] In some embodiments, the second interaction subnetwork may include multiple adaptation subnetworks and a third interaction subnetwork. Obtaining a second fusion feature using the second interaction subnetwork based on a third fusion feature and a first fusion feature may include: inputting the third fusion feature into the multiple adaptation subnetworks respectively to obtain multiple fourth fusion features accordingly, wherein the multiple adaptation subnetworks are configured to generate multiple fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and for each of the multiple fourth fusion features, inputting the fourth fusion feature and the first fusion feature into the third interaction subnetwork to obtain a fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature includes the fifth fusion feature corresponding to each of the multiple fourth fusion features.

[0148] In some embodiments, the ranking model may further include an adaptation network, and the training apparatus for the ranking model may further include: a sixth acquisition unit configured to input the first fusion feature into the adaptation network to obtain the adapted first fusion feature; and wherein the fifth acquisition unit may be further configured to: obtain the prediction score of the candidate object using a prediction network based on the second fusion feature and the adapted first fusion feature.

[0149] In some embodiments, the training apparatus for the ranking model may further include: a second determining unit configured to determine that the ranking model has completed preliminary training in response to determining that the ranking model has reached a second convergence condition based on the ranking loss training; a defixing unit configured to defix the parameters of the pre-trained first interactive network; and a third training unit configured to jointly fine-tune the networks of the ranking model based on sample data, wherein the first interactive network adopts a first learning rate, and the networks in the ranking model other than the first interactive network adopt a second learning rate, and the first learning rate is lower than the second learning rate.

[0150] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0151] According to embodiments of this disclosure, an electronic device, a readable storage medium, and a computer program product are also provided.

[0152] refer to Figure 13The present invention describes a structural block diagram of an electronic device 1300 that can serve as a server or client of the present disclosure, which is an example of a hardware device that can be applied to various aspects of the present disclosure. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0153] like Figure 13 As shown, the electronic device 1300 includes a computing unit 1301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1302 or a computer program loaded from a storage unit 1308 into a random access memory (RAM) 1303. The RAM 1303 may also store various programs and data required for the operation of the electronic device 1300. The computing unit 1301, ROM 1302, and RAM 1303 are interconnected via a bus 1304. An input / output (I / O) interface 1305 is also connected to the bus 1304.

[0154] Multiple components in electronic device 1300 are connected to I / O interface 1305, including: input unit 1306, output unit 1307, storage unit 1308, and communication unit 1309. Input unit 1306 can be any type of device capable of inputting information to electronic device 1300. Input unit 1306 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device, and may include, but is not limited to, a mouse, keyboard, touchscreen, trackpad, trackball, joystick, microphone, and / or remote control. Output unit 1307 can be any type of device capable of presenting information, and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 1308 may include, but is not limited to, a hard disk and an optical disk. The communication unit 1309 allows the electronic device 1300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers and / or chipsets, such as Bluetooth devices, 802.11 devices, WiFi devices, WiMax devices, cellular communication devices and / or the like.

[0155] The computing unit 1301 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1301 performs the various methods and processes described above, such as the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure. For example, in some embodiments, the object sorting method based on the sorting model of this disclosure or the training method of the sorting model of this disclosure can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 1308. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 1300 via ROM 1302 and / or communication unit 1309. When the computer program is loaded into RAM 1303 and executed by computing unit 1301, one or more steps of the object sorting method based on the sorting model implementation of this disclosure or the training method of the sorting model of this disclosure described above can be performed. Alternatively, in other embodiments, computing unit 1301 can be configured by any other suitable means (e.g., by means of firmware) to execute the object sorting method based on the sorting model implementation of this disclosure or the training method of the sorting model of this disclosure.

[0156] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0157] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0158] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0159] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0160] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0161] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0162] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0163] While embodiments or examples of this disclosure have been described with reference to the accompanying drawings, it should be understood that the methods, systems, and devices described above are merely exemplary embodiments or examples, and the scope of the invention is not limited by these embodiments or examples, but only by the granted claims and their equivalents. Various elements in the embodiments or examples may be omitted or replaced by their equivalents. Furthermore, the steps may be performed in a different order than that described in this disclosure. Further, various elements in the embodiments or examples may be combined in various ways. Importantly, as the technology evolves, many elements described herein can be replaced by equivalents that appear after this disclosure.

Claims

1. An object sorting method based on a sorting model, the method comprising: Obtain a candidate list, the current search term corresponding to the candidate list, and a historical behavior sequence. The historical behavior sequence includes the historical search terms of the user application and the object information of the historical clicked objects corresponding to the historical search terms. Based on the historical behavior sequence and the current search term, a first fusion feature is obtained using the first interaction network in the ranking model. The first interaction network is trained based on a generative task used to generate object information of the user's real clicked object. For each candidate object in the candidate list, a prediction score for the candidate object is obtained based on the object information of the candidate object and the first fusion feature. as well as The candidate objects in the candidate list are sorted based on the predicted score of each candidate object in the candidate list.

2. The method according to claim 1, wherein, The ranking model further includes a second interaction network and a prediction network. The step of predicting the prediction score of the candidate object based on the object information of the candidate object and the first fusion feature includes: Based on the object information of the candidate object and the first fusion feature, the second fusion feature of the candidate object is obtained using the second interaction network; and Based on the second fusion feature, the prediction network is used to obtain the prediction score of the candidate object.

3. The method according to claim 2, further comprising: Obtain the user's context information; And among them, The step of obtaining the second fusion feature of the candidate object using the second interaction network based on the object information of the candidate object and the first fusion feature includes: Based on the object information of the candidate object, the context information, and the first fusion feature, the second fusion feature of the candidate object is obtained using the second interaction network.

4. The method according to claim 3, wherein, The second interaction network includes a first interaction sub-network and a second interaction sub-network. The step of obtaining the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information, the context information, and the first fusion feature includes: The first interaction sub-network is used to perform feature interaction between the object information and the context information of the candidate object to obtain a third fused feature; and Based on the third fusion feature and the first fusion feature, the second fusion feature is obtained using the second interactive sub-network.

5. The method according to claim 4, wherein, The second interactive sub-network includes multiple adaptation sub-networks and a third interactive sub-network. The step of obtaining the second fusion feature using the second interactive sub-network based on the third fusion feature and the first fusion feature includes: The third fusion feature is input into the plurality of adaptor subnetworks to obtain a plurality of fourth fusion features, wherein the plurality of adaptor subnetworks are configured to generate the plurality of fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and For each of the plurality of fourth fusion features, the fourth fusion feature and the first fusion feature are input into the third interactive sub-network to obtain the fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature includes the fifth fusion feature corresponding to each of the plurality of fourth fusion features.

6. The method according to claim 5, wherein, The third interactive sub-network is a cross-attention network.

7. The method according to any one of claims 2 to 6, wherein, The ranking model also includes an adaptation network, and the method further includes: The first fusion feature is input into the adaptation network to obtain the adapted first fusion feature; and wherein... The step of obtaining the prediction score of the candidate object using the prediction network based on the second fusion feature includes: Based on the second fusion feature and the adapted first fusion feature, the prediction network is used to obtain the prediction score of the candidate object.

8. The method according to any one of claims 1 to 7, wherein, The object information includes a semantic identifier, and the acquisition of the semantic identifier includes: Based on object description information, a corresponding semantic vector is generated using a pre-trained large model. This pre-trained large model is obtained through fine-tuning training based on the description information of each object in an object similarity cluster, and each object similarity cluster contains multiple objects corresponding to the same search term. The semantic identifier is determined based on the semantic vector.

9. The method according to claim 8, wherein, The semantic identifiers include multi-level identifiers used to represent semantic categories from coarse to fine.

10. The method according to claim 9, wherein, Determining the semantic identifier based on the semantic vector includes: For the semantic vector, perform multiple rounds of the first operation, wherein each round of the first operation includes: Among the multiple center vectors corresponding to the current round, the target center vector that is closest to the object vector corresponding to the current round is determined, and the number corresponding to the target center vector is determined as the sub-identifier corresponding to the current round. The object vector corresponding to the first round is the semantic vector; and The residual vector between the object vector corresponding to the current round and the target center vector is determined as the object vector corresponding to the next round, where, The multiple center vectors corresponding to the current round are obtained by clustering the object vectors corresponding to the current round of each object in the object library; and The multiple sub-identifiers obtained from the first operation in each round are combined into the multi-level identifier.

11. The method according to claim 10, wherein, The object information includes multiple semantic identifiers corresponding to multiple information fields, which include at least two of the following: a search term field, an object title field, an object landing page field, and an object image field.

12. The method according to claim 11, wherein, The number of rounds of the first operation used to obtain semantic identifiers in different information domains, as well as the number of center vectors corresponding to each round of the first operation, are determined according to the semantic granularity of the corresponding information domain.

13. The method according to claim 11 or 12, wherein, After the current search term and the historical search terms are converted into semantic identifiers of the corresponding search term domains, they are input into the ranking model.

14. The method according to any one of claims 9 to 13, wherein, Each semantic identifier is processed as follows before being input into the ranking model: Each identifier in the multi-level identifier of each semantic identifier is embedded to obtain the identifier code corresponding to each identifier; as well as At least the multiple identifier codes corresponding to the multi-level identifiers are input into the merger to obtain compressed semantic identifier codes, which are then input into the sorting model.

15. The method of claim 14, wherein, The merger is built based on a Transformer network, and the step of inputting at least the multiple identifier codes corresponding to the multi-level identifiers into the merger to obtain the compressed semantic identifier codes includes: According to the hierarchical order of the identifiers in the multi-level identifiers, determine the position code corresponding to each identifier code in the multi-level identifier encoding; and The multiple identifier codes and the position codes corresponding to each identifier code are input into the merger to obtain the semantic identifier code.

16. A method for training a ranking model, the ranking model comprising a first interaction network, the method comprising: Acquire sample data, which includes sample search terms, historical behavior sequences, and object information of the actual clicked objects of users corresponding to the sample search terms; Based on the historical behavior sequence and the sample search terms, the first feature is obtained using the first interactive network; The first feature is processed using a decoder to generate predicted object information; Based on the predicted object information and the object information of the user's actual clicked object, calculate the generative loss; as well as The first interactive network is trained using the generative loss.

17. The method according to claim 16, wherein, The ranking model further includes a second interaction network and a prediction network, and the sample data further includes candidate objects corresponding to the sample search terms and labels indicating whether the candidate objects have been clicked. The method further includes: In response to the determination that the first interactive network has reached the first convergence condition based on the generative loss training, it is determined that the first interactive network has completed pre-training. Fix the parameters of the pre-trained first interaction network; Based on the historical behavior sequence and the sample search terms, the first fusion feature is obtained using the pre-trained first interaction network; Based on the object information of the candidate object and the first fusion feature, the second fusion feature of the candidate object is obtained using the second interaction network; Based on the second fusion feature, the prediction network is used to obtain the prediction score of the candidate object; The ranking model is trained based on the predicted score and the ranking loss calculated from the label.

18. The method according to claim 17, wherein, The sample data also includes the user's context information, and wherein obtaining the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information and the first fusion feature includes: Based on the object information of the candidate object, the context information, and the first fusion feature, the second fusion feature of the candidate object is obtained using the second interaction network.

19. The method of claim 18, wherein, The second interaction network includes a first interaction sub-network and a second interaction sub-network. The step of obtaining the second fusion feature of the candidate object using the second interaction network based on the candidate object's object information, the context information, and the first fusion feature includes: The first interaction sub-network is used to perform feature interaction between the object information and the context information of the candidate object to obtain a third fused feature; and Based on the third fusion feature and the first fusion feature, the second fusion feature is obtained using the second interactive sub-network.

20. The method of claim 19, wherein, The second interactive sub-network includes multiple adaptation sub-networks and a third interactive sub-network. The step of obtaining the second fusion feature using the second interactive sub-network based on the third fusion feature and the first fusion feature includes: The third fusion feature is input into the plurality of adaptor subnetworks to obtain a plurality of fourth fusion features, wherein the plurality of adaptor subnetworks are configured to generate the plurality of fourth fusion features so that the candidate object can focus on different dimensions of the first fusion feature; and For each of the plurality of fourth fusion features, the fourth fusion feature and the first fusion feature are input into the third interactive sub-network to obtain the fifth fusion feature corresponding to the fourth fusion feature, wherein the second fusion feature includes the fifth fusion feature corresponding to each of the plurality of fourth fusion features.

21. The method according to any one of claims 17 to 20, wherein, The ranking model also includes an adaptation network, and the method further includes: The first fusion feature is input into the adaptation network to obtain the adapted first fusion feature; and wherein... The step of obtaining the prediction score of the candidate object using the prediction network based on the second fusion feature includes: Based on the second fusion feature and the adapted first fusion feature, the prediction network is used to obtain the prediction score of the candidate object.

22. The method according to any one of claims 17 to 21, further comprising: In response to the determination that the ranking model has reached the second convergence condition based on the ranking loss, it is determined that the ranking model has completed the initial training. The parameters of the pre-trained first interactive network are unfixed; as well as The ranking model is jointly fine-tuned based on the sample data, wherein the first interactive network adopts a first learning rate, and all networks in the ranking model other than the first interactive network adopt a second learning rate, wherein the first learning rate is lower than the second learning rate.

23. An object sorting device based on a sorting model, the device comprising: The first acquisition unit is configured to acquire a candidate list, the current search term corresponding to the candidate list, and a historical behavior sequence, wherein the historical behavior sequence includes the historical search terms of the user application and the object information of the historical clicked objects corresponding to the historical search terms. The second acquisition unit is configured to acquire a first fusion feature based on the historical behavior sequence and the current search term, using the first interaction network in the ranking model. The first interaction network is trained based on a generative task for generating object information of the user's real clicked object. The first prediction unit is configured to predict a prediction score for each candidate object in the candidate list based on the object information of the candidate object and the first fusion feature. as well as The sorting unit is configured to sort the candidate objects in the candidate list based on the predicted score of each candidate object in the candidate list.

24. A training apparatus for a ranking model, the ranking model including a first interaction network, the apparatus comprising: The first acquisition unit is configured to acquire sample data, which includes sample search terms, historical behavior sequences, and object information of the user's actual clicked object corresponding to the sample search terms. The second acquisition unit is configured to acquire a first feature based on the historical behavior sequence and the sample search terms using the first interactive network. The generation unit is configured to process the first feature using a decoder to generate predicted object information; The first calculation unit is configured to calculate generative loss based on the predicted object information and the object information of the user's actual clicked object. as well as The first training unit is configured to train the first interactive network using the generative loss.

25. An electronic device, comprising: At least one processor; as well as A memory that is communicatively connected to the at least one processor; in The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-22.

26. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-22.

27. A computer program product comprising a computer program, wherein, When the computer program is executed by a processor, it implements the method of any one of claims 1-22.

Images