Quality evaluation model training method, information search method, information recommendation method, equipment, storage medium and program product

Abstract

The embodiment of the invention provides a quality evaluation model training method, an information search method, an information recommendation method, equipment, a storage medium and a program product. In the embodiment of the invention, for a model training process, sample selection is carried out from page dimensions, and samples in a page and negative samples outside the page are determined; constructing a double contrast loss function matched with the sample; through the double contrast loss function, model training is performed by using the in-page sample and the out-page negative sample, so that the model precision can be improved, and the accuracy of model quality evaluation is improved. Furthermore, an importance distribution function is designed for samples in the page and negative samples outside the page, importance of different negative samples in optimization is dynamically learned, difficult negative samples are mined, and model optimization is more flexibly and efficiently assisted.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a quality assessment model training method, information search method, information recommendation method, device, storage medium and program product. Background Art

[0002] In e-commerce scenarios, e-commerce applications can provide users with product search and sorting services. For example, in the scenario of image-based product search, the user enters an image on the application page, and the e-commerce application obtains multiple products similar to the image, evaluates the quality of the multiple products through the quality scoring model, and displays the products with higher scores on the application page. Users can interact with one or more products on the application page according to their own needs or preferences, such as clicking, collecting, or purchasing.

[0003] Currently, the quality rating model has the problem of low quality rating accuracy, resulting in the products displayed on the application page not meeting the user's interaction needs. Summary of the invention

[0004] Multiple aspects of the present application provide a quality assessment model training method, an information search method, an information recommendation method, a device, a storage medium and a program product to improve the assessment accuracy of the quality assessment model.

[0005] The embodiment of the present application provides a quality assessment model training method, including: obtaining at least two sample pages, the at least two sample pages including multiple sample information objects; for any sample page, identifying, from the multiple sample information objects, in-page samples belonging to any sample page and out-page negative samples that do not belong to any sample page; constructing a double contrast loss function based on the in-page samples and the out-page negative samples, the double contrast loss function being used for an initial quality assessment model to learn the user's interest in interacting with information objects and predicting the classification boundary of whether the user has interactive behavior; iteratively training the initial quality assessment model according to the double contrast loss function and the in-page samples and the out-page negative samples to obtain a target quality assessment model.

[0006] An embodiment of the present application also provides an information search method, which includes: in response to a search operation initiated by a target user, determining user information and search information corresponding to the search operation; obtaining multiple candidate information objects based on the search information; performing quality assessment on the multiple candidate information objects using a target quality scoring model based on the user information and the search information to obtain quality scores for each of the multiple candidate information objects; selecting a target information object from the multiple candidate information objects based on the quality scores of each of the multiple candidate information objects, and displaying it on a result page corresponding to the search operation; wherein the target quality scoring model is trained based on a quality assessment model training method.

[0007] An embodiment of the present application also provides an information recommendation method, including: when a set event is triggered, obtaining multiple candidate information objects; using a target quality scoring model to perform quality assessment on the multiple candidate information objects to obtain quality scores for each of the multiple candidate information objects; selecting a target information object from the multiple candidate information objects based on the quality scores of each of the multiple candidate information objects, and displaying it on a target page; wherein the target quality scoring model is trained based on a quality assessment model training method.

[0008] An embodiment of the present application also provides an electronic device, comprising: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement each step in the quality assessment model training method, information search method and information recommendation method provided in the embodiment of the present application.

[0009] An embodiment of the present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor implements each step of the quality assessment model training method, information search method, and information recommendation method provided in the embodiment of the present application.

[0010] An embodiment of the present application also provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the processor implements each step of the quality assessment model training method, information search method and information recommendation method provided in the embodiment of the present application.

[0011] In the embodiment of the present application, for the model training process, sample selection is performed from the page dimension to determine the samples within the page and the negative samples outside the page; and a dual contrast loss function adapted to the sample is constructed; through the dual contrast loss function, the model is trained using the samples within the page and the negative samples outside the page, which can improve the model accuracy and the accuracy of the model quality assessment. Furthermore, an importance distribution function is designed for the samples within the page and the negative samples outside the page, and the importance of different negative samples in optimization is dynamically learned, and "difficult negative samples" are mined, which can assist in model optimization in a more flexible and efficient manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0013] Figure 1 A flowchart of a quality assessment model training method provided for an exemplary embodiment of the present application;

[0014] Figure 2aA schematic diagram of a sample page provided for an exemplary embodiment of the present application;

[0015] Figure 2b A schematic diagram of the relationship between a sample page and an information object provided for an exemplary embodiment of the present application;

[0016] Figure 2c A schematic diagram of a quality assessment model training and reasoning structure provided for an exemplary embodiment of the present application;

[0017] Figure 3 A schematic diagram of the structure of a quality assessment model training device provided for an exemplary embodiment of the present application;

[0018] Figure 4 A schematic structural diagram of an electronic device provided for an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0020] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0021] The various models involved in this application (including but not limited to language models or large models) are in compliance with relevant laws and standards.

[0022] Considering that the user's interactive behavior on the application page is highly dependent on the products displayed on the page. For example, when a user interacts with some of the products on the application page, it can be considered that these products have certain specificities, while other products on the application page do not have this specificity, so that the user chooses to interact with these selected products instead of interacting with other products on the application page. Therefore, understanding and modeling the user's interactive behavior towards products on different application pages can enable the quality assessment model to more accurately mine the user's personalized interaction mode, improve the accuracy of the quality assessment model in quality assessment, and promote the user's interaction with the products on the application page.

[0023] One way to improve the accuracy of the quality assessment model through context modeling is as follows: add an encoder, such as a graph neural network (GNN), before the quality assessment model, input the target product and its context information on the application page that interacts with the user into the encoder for encoding, and input the encoded features into the quality assessment model for quality assessment. The context information refers to other products on the application page except the target product.

[0024] The problems with the above method are:

[0025] 1) It is necessary to introduce an additional encoder (such as GNN) before the quality assessment model, which increases the computational complexity of the entire process, especially during model inference, which increases the loss of computing resources and the delay.

[0026] 2) When providing online real-time reasoning services, it is difficult to obtain the target product and its context information on the application page, because, under normal circumstances, the target product and its context information are determined after the application page is displayed, and before the application page is displayed, the user cannot perform interactive operations and cannot determine the target product and its context information. Therefore, it is necessary to determine approximate context information through a specific algorithm, for example, using multiple candidate products obtained based on search information as target products and their context information, which may lead to inconsistent distribution of input information during training and reasoning. Therefore, the effect of online real-time reasoning is far inferior to that of offline evaluation.

[0027] In the embodiment of the present application, from the perspective of the model training process of the quality assessment model, for the model training process, sample selection is performed from the page dimension to determine the samples within the page and the negative samples outside the page; and a dual contrast loss function adapted to the sample is constructed; through the dual contrast loss function, the model is trained using the samples within the page and the negative samples outside the page, which can improve the model accuracy and the accuracy of the model quality assessment, making it possible for the quality assessment model to be independent of the pre-model. Furthermore, an importance distribution function is designed for the samples within the page and the negative samples outside the page, and the importance of different negative samples in optimization is dynamically learned to mine "difficult negative samples" for more flexible and efficient auxiliary model optimization. In addition, since the dual contrast loss function does not participate in model reasoning, therefore, when the model is inferred online, the loss of computing resources is not increased, and the pre-encoder is no longer relied on. Since it no longer relies on the pre-encoding network, the problems of loss of computing resources, large delay, and low accuracy caused by the introduction of the encoder can be solved.

[0028] A solution provided by an embodiment of the present application is described in detail below in conjunction with the accompanying drawings.

[0029] Figure 1 A flow chart of a quality assessment model training method provided for an exemplary embodiment of the present application is as follows: Figure 1 As shown, the method includes:

[0030] 101. Obtain at least two sample pages, where the at least two sample pages include a plurality of sample information objects;

[0031] 102. For any sample page, identify, from a plurality of sample information objects, in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page;

[0032] 103. Construct a double contrast loss function based on the samples in the page and the negative samples outside the page. The double contrast loss function is used for the initial quality assessment model to learn the user's interest in the interaction of the information object and to predict the classification boundary of whether the user has an interactive behavior;

[0033] 104. According to the dual contrast loss function and the on-page samples and off-page negative samples, the initial quality assessment model is iteratively trained to obtain the target quality assessment model.

[0034] In this embodiment, the quality assessment model is a model that provides quality assessment services for information objects for applications. The implementation method of the quality assessment model is not limited, and any model that can perform quality assessment on information objects is applicable to the embodiments of this application. For example, the quality assessment model may include, but is not limited to: a wide-depth (WideDeep) model and a DeepFM model. Among them, the WideDeep model improves the expressiveness and generalization ability of the model through the advantages of shallow models (such as linear models) and deep models (such as neural networks). Among them, the DeepFM model is a model that combines a factorization machine (Factorization Machine, FM) and a deep neural network (DNN) to achieve efficient feature interaction capture.

[0035] The quality assessment model has different functions in different application scenarios, mainly including quality scoring, sorting, and determining the classification boundaries of information objects. An exemplary description is given below.

[0036] (1) Quality scoring. The quality assessment model scores information objects (such as products) and gives quantitative indicators to represent the quality of the information object. For example, the quality assessment model is used to score the compatibility between users and products. A high score indicates that the product is highly compatible with the user and is suitable for recommendation to the user. The quality assessment model can give a specific score value for the information object, for example, 0-10 points. The higher the score, the higher the compatibility between the information object and the user.

[0037] (2) Sorting. Focus on the relative order between information objects, give relative scores between multiple information objects, and sort multiple information objects based on the relative scores so that users can see information objects with high adaptability or high relevance to users.

[0038] (3) Determine the classification boundary of information objects. The quality assessment model can learn which feature combinations are highly compatible with users and which feature combinations are less compatible with users, so as to automatically determine which information objects can be displayed to users and which cannot be displayed to users. For example, the quality assessment model can not only output the quality scores of multiple information objects, but also give a score threshold and display information objects with quality scores higher than the score threshold to users.

[0039] The application may include but is not limited to: shopping applications, social applications, online learning applications, and search engine applications. The application corresponds to an application page, and the application page displays information objects. The information objects on the application page may be carried in the form of cards, pictures, links, and texts.

[0040] Depending on the application, the information objects displayed on the application page are different. If the application is a shopping application, the information objects displayed on the application page may be product objects; if the application is a social application, the information objects displayed on the application page are social information, such as posts published by users; if the application is a search engine application, the information objects displayed on the application page are search information, such as news, web links, and advertisements; if the application is an online education application, the information objects displayed on the application page are educational information, such as courses, lectures, and quizzes.

[0041] In this embodiment, for the convenience of distinction and description, the application page used for model training is called a sample page, and the information object included in the sample page is called a sample information object. Among them, at least two sample pages are used as sample data for model training of the initial quality assessment model, and at least two sample pages include multiple sample information objects. Here, multiple sample information objects refer to a collection of sample information objects contained in at least two sample pages. For example, taking two sample pages as an example, sample page A1 includes 6 information objects, and sample page A2 includes 5 information objects, then sample page A1 and sample page A2 have 11 corresponding information objects. Figure 2aAs shown, a search scenario of an e-commerce application is used as an example. The user's search request is "Coke". Six products that match "Coke" are displayed on the sample page D1. The solid-line frame represents the product clicked by the user, and the dotted-line frame represents the product not clicked by the user. In addition to the sample page D1, another sample page D2 is provided, which includes six products.

[0042] In this embodiment, for any sample page, the in-page samples belonging to the any sample page and the off-page negative samples not belonging to the any sample page are identified from the multiple sample information objects. Wherein, for any sample page, the in-page samples refer to the sample information objects included in the any sample page; the off-page samples refer to the sample information objects included in other sample pages except the any sample page. Optionally, the in-page samples include in-page positive samples and in-page negative samples. Wherein, the in-page positive samples refer to the sample information objects among the multiple sample information objects that belong to the any sample page and interact with the user, and the in-page negative samples refer to the sample information objects among the multiple sample information objects that belong to the any sample page and do not interact with the user. The off-page samples include off-page positive samples and off-page negative samples. Wherein, the off-page positive samples refer to the sample information objects among the multiple sample information objects that do not belong to the any sample page and interact with the user, and the off-page negative samples refer to the sample information objects among the multiple sample information objects that do not belong to the any sample page and do not interact with the user. Combined with Figure 2a , relative to sample page D1, the 6 products in sample page D1 are on-page samples, and the information objects in the solid-line box are on-page positive samples, and the information objects in the dotted-line box are on-page negative samples; relative to sample page D1, the 6 products in sample page D2 are all off-page samples, and the off-page positive samples are not illustrated. Figure 2a The sample pages and information objects in the examples are only examples and are not limited to these.

[0043] Among them, the interactive behavior with the user means that the user performs an interactive behavior operation on the sample information object. Depending on the application, the interactive behavior operations performed by the user on the sample information object are also different. For example, if the application is a shopping application, the interactive behavior operations performed by the user on the sample information object may include but are not limited to: purchase, collection, click to view, add to purchase, and share. If the application is a social application, the interactive behavior operations performed by the user on the sample information object may include but are not limited to: like, comment, forward, and share. If the application is a search engine application, the interactive behavior operations performed by the user on the sample information object may include but are not limited to: click to view, share, collect, and comment. If the application is an online education application, the interactive behavior operations performed by the user on the sample information object may include but are not limited to: listening to classes, subscribing, participating in tests, and collecting.

[0044] In this embodiment, a dual contrast loss function is constructed based on the on-page samples and the off-page negative samples. The dual contrast loss function can compare the user's interactive behaviors on different sample pages through the two dimensions of on-page and off-page, and can better capture the preference differences of users in different application pages (also known as different contexts), thereby improving the accuracy of the quality assessment of the model. In short, the dual contrast loss function involves the contrast loss on the page and the contrast loss on the off-page.

[0045] Among them, the contrast loss within the page describes the comparison between multiple sample information objects within the same sample page. For example, a user may select a target sample information object of interest from multiple similar sample information objects (such as commodities) and perform interactive behavior operations on the target sample information object. Some features of the multiple similar sample information objects are very similar. The initial quality assessment model can learn how to distinguish the sample information object of interest to the user from multiple similar sample information objects.

[0046] Among them, the off-page contrast loss describes the comparison of sample information objects of different sample pages. By introducing diverse negative samples (i.e., sample information objects that users are not interested in), the model can better learn which feature combinations users are really interested in, which helps the model learn classification boundaries more accurately and avoid overfitting.

[0047] Through the above analysis, it can be known that the dual contrast loss function in the embodiment of the present application is used for the initial quality assessment model to learn the user's interest in interacting with information objects and predict the classification boundaries of whether the user has interactive behavior. The in-page contrast loss is mainly used to assist the initial quality assessment model in learning the user's interest preferences for multiple similar information objects in the application page, and the out-page contrast loss is mainly used to provide the model with more diverse negative samples to assist the model in better learning the classification boundaries. Of course, the in-page contrast loss can also assist in learning the classification boundaries, and the out-page contrast loss can also assist in learning the user's interest preferences.

[0048] The classification boundary refers to the decision surface or rule learned by the initial quality assessment model in the feature space to distinguish between positive samples (sample information objects that users are interested in) and negative samples (sample information objects that users are not interested in). By considering the comparison within the page and the comparison outside the page, the dual contrast loss function can help the model better learn this classification boundary, thereby improving the accuracy and diversity of the initial quality assessment model. The comparison within the page can ensure that the model can make accurate choices between similar sample information objects, while the comparison outside the page provides more diverse negative samples, helping the model better generalize to unseen data.

[0049] Specifically, the classification boundary is used to predict whether the user interacts with the information object. The classification boundary can define which feature combinations can make the user interact with the information object and which feature combinations will not interact with the user in the feature space corresponding to the quality assessment model. Among them, the features in the feature space refer to the features of the information object.

[0050] Among them, the characteristics of information objects vary depending on the type of application. For example, for shopping applications, the characteristics of information objects include but are not limited to: name, price, place of shipment, user evaluation, brand, discount information, sales volume and brand, etc. For social applications, the characteristics of information objects include but are not limited to: user name, avatar, release time, number of comments, number of shares, tags, topics and geographic location, etc. For search engine applications, the characteristics of information objects may include but are not limited to: title, link, summary, click-through rate, language, source website and release date, etc. For online learning applications, the characteristics of information objects include but are not limited to: course name, course description, instructor information, document materials, video materials, homework, learning progress, evaluation and certificate, etc.

[0051] In this embodiment, the initial quality assessment model is iteratively trained according to the dual contrast loss function and the samples within the page and the negative samples outside the page to obtain the target quality assessment model. During the iterative training process, the initial quality assessment model can be gradient backpropagated and parameter optimized through the dual contrast loss function, and the dual contrast loss function is recalculated through the initial quality assessment model after parameter optimization until the set iteration stop condition is met. Among them, the set iteration stop condition includes: reaching the set number of iterations, reaching the set iteration time, the dual contrast loss function value no longer decreases, or the dual contrast loss function value reaches the set threshold, etc.

[0052] In the embodiment of the present application, for the process of iterative model training, sample selection is performed from the page dimension to determine the samples within the page and the negative samples outside the page; and a dual contrast loss function adapted to the samples is constructed; through the dual contrast loss function, the model is trained using the samples within the page and the negative samples outside the page, which can improve the model accuracy and the accuracy of the model quality assessment. Furthermore, an importance distribution function is designed for the samples within the page and the negative samples outside the page, and the importance of different negative samples in optimization is dynamically learned, and "difficult negative samples" are mined, which is more flexible and efficient in assisting model optimization.

[0053] In an optional embodiment, each sample page corresponds to page identification information, and the page identification information can be any information that can uniquely identify the sample page. For example, the page identification information can include, but is not limited to: a session identifier (ID), a uniform resource locator (URL), a page title or a page description, etc. Each information object corresponds to behavior tag information, and the behavior tag information is used to identify whether the sample information object has user interaction data. Among them, when a user performs an interactive behavior operation on a sample information object, user interaction data can be saved for the sample information object, and the user interaction data may include, but is not limited to: descriptive information such as the type, name, time, and content of the user interaction operation. If the user does not perform an interactive behavior operation on the sample information object, the sample information object does not have user interaction data.

[0054] The sample data can be organized in advance according to the requirements of the dual contrast loss function. The sample data includes page identification information, behavior label information, and features corresponding to the information object. The sample features can refer to the above introduction to the features of the information object. The information objects with the same page identification information are integrated together to represent the information objects under the same search request. Figure 2bAs shown in the figure, the page identification information is implemented as a session ID, and the behavior tag information is illustrated as 0 or 1. The behavior tag information is 0, which indicates negative feedback, that is, no user interaction data, and the behavior tag information is 1, which indicates positive feedback, that is, there is user interaction data. The first session ID1 is used as a reference, which is the positive sample in the page. Figure 2b The sample information object with session ID 1 and behavior tag information 0 is taken as the negative sample in the page, and the sample information object with session ID 2 and behavior tag information 0 is taken as the negative sample outside the page.

[0055] Among them, Figure 2a The products in the solid-line frame on the sample page shown in the figure are positive samples within the page, the products in the dotted-line frame on the sample page are negative samples within the page, and the products in the dotted-line frames on other sample pages are negative samples outside the page.

[0056] Based on the above, an implementation method for identifying, for any sample page, in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page from multiple sample information objects includes: for any sample page, identifying, from multiple sample information objects, in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page according to page identification information and behavior label information corresponding to multiple sample information objects. The sample information objects located on the same sample page correspond to the same page identification information.

[0057] Optionally, the behavior label information includes positive feedback behavior label information and negative feedback behavior label information. Positive feedback behavior label information indicates that the sample information object has user interaction data, and negative feedback behavior label information indicates that the sample information object does not have user interaction data. For the convenience of distinction and description, the sample information object with positive feedback behavior label information is called a positive sample within the page, and the sample information object with negative feedback behavior label information is called a negative sample within the page.

[0058] On this basis, an implementation method is provided for identifying in-page samples belonging to any sample page and off-page negative samples not belonging to any sample page from multiple sample information objects according to page identification information and behavior label information corresponding to the multiple sample information objects, including: dividing the multiple sample information objects into a first sample information object set belonging to any sample page and a second sample information object set not belonging to any sample page according to the page identification information corresponding to the multiple sample information objects; taking the sample information objects in the first sample information object set as in-page samples belonging to any sample page, the in-page samples including in-page positive samples with positive feedback behavior label information and in-page negative samples with negative feedback behavior label information; taking the negative sample information objects with negative feedback behavior label information in the second sample information object set as off-page negative samples belonging to any sample page.

[0059] In an optional embodiment, an implementation method of constructing a dual contrast loss function based on on-page samples and off-page negative samples includes: for any sample page, determining the on-page quality score variable corresponding to the on-page sample and the off-page quality score variable corresponding to the off-page negative sample; and constructing a dual contrast loss function of an initial quality assessment model based on the on-page quality score variables and the off-page quality score variables of at least two sample pages.

[0060] Optionally, the samples on the page include: positive samples on the page and negative samples on the page. An implementation method for determining the quality score variables on the page corresponding to the samples on the page and the quality score variables off the page corresponding to the negative samples off the page for any sample page, including: for any sample page, determining the first quality score variable corresponding to the positive sample on the page, the second quality score variable corresponding to the negative sample on the page, and the third quality score variable corresponding to the negative sample off the page; An implementation method for constructing a double contrast loss function of an initial quality assessment model based on the quality score variables on the page and the quality score variables off the page of at least two sample pages, including: constructing a double contrast loss function of an initial quality assessment model based on the first quality score variable, the second quality score variable and the third quality score variable of at least two sample pages. Wherein, in the case of calculating the double contrast loss value later, the first quality score of the positive sample on the page can be assigned to the first quality score variable, the second quality score of the negative sample on the page can be assigned to the second quality score variable, and the third quality score of the negative sample off the page can be assigned to the third quality score variable.

[0061] The following is an exemplary description of the dual contrast loss function, but is not limited thereto.

[0062] For any sample page, the positive sample in any page is recorded as v +, the set of negative samples in the sample page is called V I - , the off-page negative sample set corresponding to the sample page is V O - . Positive samples in the page v + The first quality score variable is denoted as s + , the off-page negative sample set V I - The quality score variable is recorded as Off-page negative sample set V O - The quality score variable is recorded as A method for positive samples v in any page + An example of the sub-double contrast loss function is expressed as follows:

[0063]

[0064] Where E is the mathematical expectation, τ is the heat coefficient, τ takes a value between 0 and 1, for example, 0.07, 0.1 or 0.5, and ∪ is the union. + is an example of the first scoring variable, is an example of a second quality score variable set, wherein the second quality score variable set includes one or more second quality score variables. is an example of a third quality score variable set, and the third quality score variable set includes one or more third quality score variables.

[0065] Accordingly, the dual contrast loss function corresponding to the initial quality assessment model can be constructed according to the sub-dual contrast loss functions corresponding to the positive samples in the page included in multiple sample information objects. For example, weighted addition is performed on the sub-dual contrast loss functions corresponding to the positive samples in the page to obtain the dual contrast loss function corresponding to the initial quality assessment model. For example, when the weight is 0.5, the sub-dual contrast loss functions corresponding to the positive samples in the page are averaged to obtain the dual contrast loss function corresponding to the initial quality assessment model. In this embodiment, the initial quality assessment model is iteratively trained according to the dual contrast loss function and the samples in the page and the negative samples outside the page to obtain the target quality assessment model. The implementation method is not limited.

[0066] In an optional embodiment, an implementation method of iteratively training an initial quality assessment model based on a dual contrast loss function and in-page samples and out-page negative samples to obtain a target quality assessment model includes: for any sample page, using the initial quality assessment model to perform quality assessment on the in-page samples and out-page negative samples of any sample page respectively to obtain an in-page quality score and an out-page quality score corresponding to any sample page; substituting the in-page quality scores and out-page quality scores corresponding to at least two sample pages into the dual contrast loss function to obtain a dual contrast loss function value; when the dual contrast loss function value does not meet the iteration stop condition, continuing to optimize the parameters of the initial quality assessment model until the iteration stop condition is met to obtain the target quality assessment model.

[0067] Optionally, any sample page also includes: sample user information and sample retrieval information. The sample user information refers to the user who requires the sample page, and the sample user information refers to information related to the sample user, such as user number, geographic location, historical behavior, device information, and user portrait. Sample retrieval information refers to specific retrieval information provided in a retrieval scenario. For example, the retrieval information may include but is not limited to: text or images. The retrieval scenario may include a user-initiated search scenario, in which case the sample retrieval information may be a search term or retrieval image entered by the user. The retrieval scenario may also include information object scenarios actively promoted by an application, such as opening an e-commerce application and the e-commerce application actively displays products on the home page. In this case, the sample retrieval information may be the user's historical browsing information and interest preference information. Based on this, an implementation method of using an initial quality assessment model to perform quality assessment on the in-page samples and off-page negative samples of any sample page respectively, so as to obtain the in-page quality score and off-page quality score corresponding to any sample page, includes: inputting the sample user information, sample retrieval information and in-page samples of any sample page into the initial quality assessment model for quality assessment, and obtaining the in-page quality score corresponding to any sample page; inputting the sample user information, sample retrieval information and off-page negative samples of any sample page into the initial quality assessment model for quality assessment, and obtaining the off-page quality score corresponding to any sample page.

[0068] Further optionally, the sample user information, sample retrieval information and positive samples within any sample page can be input into the initial quality assessment model for quality assessment to obtain a first quality score corresponding to the positive samples within the page; the sample user information, sample retrieval information and negative samples within any sample page can be input into the initial quality assessment model for quality assessment to obtain a second quality score corresponding to the positive samples within the page; the quality score outside the page corresponding to any sample page is used as the third quality score.

[0069] Optionally, the quality score within the page includes: a first quality score set adapted to the positive samples within the page included in the samples within the page and a second quality score set adapted to the negative samples within the page included in the samples within the page; the quality score outside the page includes a third quality score set adapted to the negative samples outside the page of any sample page. Among them, any sample page includes one or more positive samples within the page, any sample page includes one or more negative samples within the page, and any sample page corresponds to one or more negative samples outside the page. Accordingly, the first quality score set includes the first quality scores corresponding to the one or more positive samples within the page, the second quality score set includes the second quality scores corresponding to the one or more negative samples within the page, and the third quality score set includes the third quality scores corresponding to the one or more negative samples outside the page.

[0070] Based on this, an implementation method of substituting the on-page quality scores and off-page quality scores corresponding to at least two sample pages into a double contrast loss function to obtain a double contrast loss function value includes: for any on-page positive sample corresponding to any sample page, according to the first quality score, the second quality score set and the third quality score set corresponding to the any on-page positive sample in the first quality score set, calculating the sub-double contrast loss value of the on-page positive sample; generating the double contrast loss function value corresponding to the initial quality assessment model according to the sub-double contrast loss value of the on-page positive sample included in the at least two sample pages.

[0071] Optionally, the importance of different on-page samples and off-page negative samples in the optimization process is also different. For example, optimizing "difficult" negative samples that are closer to the classification boundary helps the model converge quickly and the classification boundary is more robust. Among them, "difficult" negative samples refer to sample information objects that do not have user interaction behavior data and have a high similarity with positive samples. Therefore, in the dual contrast loss function, an importance distribution is designed for on-page negative samples and off-page negative samples, and the importance of different negative samples is automatically adjusted during the optimization process, and "difficult negative samples" are mined more flexibly. An example of an importance distribution function is as follows:

[0072]

[0073] Among them, w s- This is an example of weight information of negative samples within a page or weight information of negative samples outside a page, but is not limited thereto.

[0074] From the above formula, we can see that when the score of the negative sample is s - The larger the value, the greater the weight w s-The larger the value, the more likely it is to be. This will prompt the model to focus on optimizing these "hard negative samples" with larger weights. For example, in an e-commerce application scenario, a user wants to buy product A. The e-commerce application provides purchase links B1, B2, and B3 corresponding to product A on the application page. The user clicks on purchase link B1 and buys product A provided by link B1. Among them, purchase link B2 and purchase link B3 are negative samples that the user did not click on, and purchase link B2 and purchase link B3 both have a high degree of similarity with purchase link B1, that is, they have more similar features. Therefore, purchase link B2 and purchase link B3 are "hard negative samples". Therefore, it is hoped that from purchase link B2 and purchase link B3, we can learn what caused the user to click on purchase link B1 instead of purchase link B2 and purchase link B3. For example, the reason may be information such as the place of shipment, delivery speed, and promotions.

[0075] Based on this, an implementation method of a sub-double contrast loss value of a positive sample in any page according to the first quality score, the second quality score set and the third quality score set corresponding to the positive sample in any page in the first quality score set includes: determining the weight information corresponding to the negative sample in the page and the weight information corresponding to the negative sample outside the page according to the second quality score set and the third quality score set corresponding to any sample page; determining the negative sample quality score according to the second quality score set and its corresponding weight information and the third quality score set and its corresponding weight information; generating the sub-double contrast loss value corresponding to the positive sample in any page according to the negative sample quality score and the first quality score corresponding to the positive sample in any page in the first quality score set.

[0076] Among them, after the importance distribution is corrected, an example of a sub-double contrast loss function corresponding to a positive sample in a page is expressed as the following formula (3):

[0077]

[0078] Among them, M is the total number of negative samples on and off the page, is the quality score of negative samples, s + is the first quality score corresponding to the positive sample in any page in the first quality assessment set.

[0079] As can be seen from the above, the dual contrast loss function considers the contextual contrast between off-page negative samples and on-page negative samples. The off-page negative sample contrast expands the model to more diverse negative sample types; the on-page negative sample contrast helps the model understand the reasons why users interact with only some information objects among many related information objects (such as differences in price, sales volume, store level, etc.). The importance distribution function corresponding to formula (2) can dynamically learn the importance of off-page negative samples and on-page negative samples in the optimization process, helping the model converge faster and reach the optimal solution.

[0080] The advantages of this dual contrast loss function over the traditional binary classification loss function used for quality assessment models are:

[0081] 1) The traditional binary classification loss function only considers the classification boundary and ignores the structural distribution of samples within the class, that is, the correlation between positive samples and negative samples. The dual contrast loss function proposed in this application can take into account the relative importance between negative samples within the page and negative samples outside the page;

[0082] 2) The dual contrast loss function proposed in this application can introduce more diverse negative samples (off-page negative samples) to alleviate the modeling bias problem caused by exposure bias. Exposure bias refers to the problem of uneven data distribution caused by the exposure mechanism in historical data, which in turn affects model training. Specifically, users are more likely to click or purchase products ranked at the top, while products ranked at the bottom may not be noticed by users due to insufficient exposure even if they are of high quality. This bias will cause the model to rely too much on historical ranking results during training, making it impossible to accurately learn the true quality of the product.

[0083] In the above embodiment, the loss function of the initial quality assessment model is a double contrast loss function for explanation. In an optional embodiment, the double contrast loss function is used to train the initial quality assessment model, which can be applied in a sorting scenario, that is, to determine the relative order of multiple information objects without considering the precise score of each information object. The following is an example in which the loss function of the initial quality assessment model is a double contrast loss function and a binary classification loss function. In an optional embodiment, considering that the binary classification loss function can enable the initial quality assessment model to output a more accurate quality score, the double contrast loss function and the binary classification loss function are used to train the initial quality assessment model, which can be applied to application scenarios focusing on sorting and quality scoring.

[0084] In another optional embodiment, an implementation method of iteratively training the initial quality assessment model according to the dual contrast loss function and the in-page samples and the off-page negative samples to obtain a target quality assessment model includes: obtaining a pre-constructed binary classification loss function, which is defined according to multiple quality scoring variables corresponding to multiple sample information objects; taking the dual contrast loss function and the binary classification loss function as optimization targets, and at least using the in-page samples and the off-page negative samples to train the initial quality assessment model to obtain the target quality assessment model.

[0085] Among them, multiple sample information objects can be input into the initial quality assessment model for quality assessment to obtain multiple quality scores corresponding to the multiple sample information objects. Alternatively, in the aforementioned embodiment, the quality score of the sample in the page is calculated for any sample page, and the multiple quality scores of the samples in the page corresponding to at least two sample pages are used as the multiple quality scores corresponding to the multiple sample information objects.

[0086] Among them, the above method is to jointly optimize the initial quality assessment model through the double contrast loss function and the binary classification loss function, and the loss function value of the initial quality assessment model can be determined based on the double contrast loss function and the binary classification loss function. The initial quality assessment model is gradient backpropagated and parameter optimized through the loss function value, and the double contrast loss value and the binary classification loss value are determined based on the initial quality assessment model after parameter optimization until the iteration stop condition is met to obtain the target quality assessment model.

[0087] Optionally, an implementation method of using a dual contrast loss function and a binary classification loss function as optimization targets, and using at least in-page samples and off-page negative samples to train an initial quality assessment model to obtain a target quality assessment model includes: using the initial quality assessment model to perform quality scores on multiple sample information objects; substituting the quality scores of the multiple sample information objects into the binary classification loss function to obtain the binary classification loss function value; for any sample page, using the initial quality assessment model to perform quality assessment on the in-page samples and off-page negative samples of any sample page respectively to obtain the in-page quality score and the off-page quality score corresponding to any sample page; substituting the in-page quality scores and the off-page quality scores corresponding to at least two sample pages into the dual contrast loss function to obtain the dual contrast loss function value; when at least one of the binary classification loss function value and the dual contrast loss function value does not meet the corresponding iteration stop condition, continuing to optimize the parameters of the initial quality assessment model until the iteration stop condition is met to obtain the target quality assessment model. Among them, the iteration stopping conditions include: the binary classification loss function value is lower than the set first loss threshold and the double contrast loss function value is lower than the set second loss threshold, the model training times reaches the set number of iterations, the model training time reaches the set training time, or the change in the binary classification loss function value is less than the set first change threshold and the change in the double contrast loss function value is less than the set second change threshold.

[0088] The first loss threshold may be 0.1 or 0.05, etc., and the second loss threshold may be 0.01 or 0.02, etc. The set number of iterations may be 1,000, 10,000, or 100,000, etc., and the set iteration time may be 2 seconds, 10 minutes, or 24 hours, etc. The first change threshold may be 0.001 or 0.002, and the second change threshold may be 0.002 or 0.003, etc.

[0089] like Figure 2c As shown, the initial quality assessment model is jointly optimized by a binary classification loss function and a dual contrast loss function as an example, but is not limited to this.

[0090] In the embodiment of the present application, the advantages of the dual modeling loss function over the existing context information modeling (e.g., adding an encoder before the quality assessment model) are:

[0091] (1) There is no need to design a specific network structure to encode context information, and online reasoning does not increase complexity.

[0092] The existing context modeling method to improve the accuracy of quality assessment models relies on user interaction behavior in real-time streams, and requires approximate acquisition during online reasoning. It has high dependence on real-time streams, increases coding losses when the model is online, and has high requirements for real-time links.

[0093] In the embodiment of the present application, the dual contrast loss function is used to optimize the model parameters when modeling context information, which does not involve changes to the structure of the initial quality assessment model itself, and does not require a specific network structure to encode context information. The dual contrast loss function takes effect during the model training process, but not during the online reasoning process, so it does not increase the complexity of online reasoning.

[0094] (2) Since the model structure of the initial quality assessment model itself is not changed, the initial quality assessment model can adopt any model architecture that is compatible with the application scenario requirements, thereby improving the versatility of the initial quality assessment model.

[0095] (3) There is no need to design specific algorithms to approximate the features of context information, thus avoiding inconsistency between training and inference feature distributions.

[0096] In the embodiment of the present application, there is no need to input context information, so online service reasoning does not rely on the context information, and there is no need to design a specific algorithm to approximate the features of the context information, which increases the accuracy of online real-time reasoning of the quality assessment model.

[0097] In addition to providing the above-mentioned quality assessment model training method, the embodiment of the present application also provides an information search method, which includes:

[0098] In response to a search operation initiated by a target user, determining user information and search information corresponding to the search operation;

[0099] According to the search information, multiple candidate information objects are obtained;

[0100] According to the user information and the search information, the target quality scoring model is used to evaluate the quality of the multiple candidate information objects to obtain the quality scores of the multiple candidate information objects;

[0101] Selecting a target information object from the multiple candidate information objects according to their respective quality scores, and displaying the target information object on a result page corresponding to the search operation;

[0102] Among them, the target quality scoring model is trained based on the above-mentioned quality assessment model training method.

[0103] In this embodiment, the information search method can be implemented on a terminal device, or it can be implemented by the cooperation of a terminal device and a server device. In the case where the information search method is implemented by the cooperation of a terminal device and a server device, the terminal device responds to a search operation initiated by a target user, determines the user information and search information corresponding to the search operation; provides the user information and search information to the server device, and the server device obtains multiple candidate information objects based on the search information; based on the user information and search information, the target quality scoring model is used to evaluate the quality of the multiple candidate information objects to obtain the quality scores of each of the multiple candidate information objects; based on the quality scores of each of the multiple candidate information objects, a target information object is selected from the multiple candidate information objects, and the target information object is provided to the terminal device; the terminal device displays the target information object on the result page corresponding to the search operation.

[0104] In this embodiment, the information search method corresponds to the scenario where the user actively initiates the search, and the user information may include but is not limited to: user ID, geographic location, historical behavior, device information, user portrait, etc. The search information may be text information or image information.

[0105] In this embodiment, the implementation method of obtaining multiple candidate information objects according to the search information is not limited. For example, the search information can be matched with the information objects in the information object library by similarity, and the information objects with a matching degree higher than a set similarity threshold are used as candidate information objects, where the similarity threshold can be 80%, 90%, 95%, etc.

[0106] In this embodiment, based on user information and search information, a target quality scoring model is used to perform quality assessment on multiple candidate information objects to obtain the quality scores of each of the multiple candidate information objects. The implementation method may be to input user information, search information, and multiple candidate information objects into the target quality scoring model for quality assessment to obtain the quality scores of each of the multiple candidate information objects.

[0107] In this embodiment, the implementation method of selecting a target information object from a plurality of candidate information objects according to the quality scores of each of the plurality of candidate information objects is not limited. For example, the number of display positions on the result page is N, N is a positive integer, and N ≥ 1; according to the number of display positions N, in the order of the quality scores of the plurality of candidate information objects from high to low, N candidate information objects are selected as target information objects. For another example, a quality score threshold is set, and candidate information objects with quality scores greater than the score threshold are selected as target information objects.

[0108] The present application also provides an information recommendation method, which includes:

[0109] When a setting event is triggered, multiple candidate information objects are obtained;

[0110] Using the target quality scoring model to evaluate the quality of multiple candidate information objects, so as to obtain the quality scores of each of the multiple candidate information objects;

[0111] Selecting a target information object from the multiple candidate information objects according to their respective quality scores, and displaying the target information object on the target page;

[0112] Among them, the target quality scoring model is trained based on the above-mentioned quality assessment model training method.

[0113] In this embodiment, the information recommendation method can be implemented on a terminal device, or it can be implemented by the cooperation of a terminal device and a server device. In the case where the information recommendation method is implemented by the cooperation of a terminal device and a server device, when a set event is triggered, the terminal device sends information recommendation demand information to the server device; the server device determines multiple candidate information objects based on the information recommendation demand information, and uses a target quality scoring model to perform quality assessment on the multiple candidate information objects to obtain quality scores for each of the multiple candidate information objects; based on the quality scores of the multiple candidate information objects, a target information object is selected from the multiple candidate information objects, and the target information object is returned to the terminal device; the terminal device displays the target information object on the target page.

[0114] In this embodiment, the information recommendation method corresponds to a scenario of actively recommending information objects to users. When a set event is triggered, the execution of the information recommendation method is triggered. Among them, the set event may include but is not limited to: application events, user operations or system events, etc. Among them, user operations include but are not limited to: various interaction methods such as touch operations, gesture operations, voice operations, head movement operations, eye movement operations, etc. Touch operations include but are not limited to: stand-alone operations, double-click operations, long press operations, and sliding operations, etc. Application events may include but are not limited to: application startup events, jump events or clock events from other applications to target applications, etc. System events may include but are not limited to: power-on events, events to recover from sleep mode, and hardware connection events, etc.

[0115] In this embodiment, when a set event is triggered, multiple candidate information objects are obtained. The multiple candidate information objects can be matched from the information object library according to user information. The user information may include but is not limited to: user number, geographic location, historical behavior, device information, and user portrait.

[0116] In this embodiment, the implementation method of using the target quality scoring model to evaluate the quality of multiple candidate information objects to obtain the quality scores of each of the multiple candidate information objects, and the implementation method of selecting the target information object from the multiple candidate information objects according to the quality scores of each of the multiple candidate information objects can be referred to the above embodiments, which will not be repeated here. Among them, the target page can be the home page of the application or other pages, which is not limited to this.

[0117] Regarding the embodiments of the present application Figure 1 The detailed implementation and beneficial effects of each step in the method have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0118] It should be noted that the execution subject of each step of the method provided in the above embodiment can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 101 to 103 can be device A; for another example, the execution subject of steps 101 and 102 can be device A, and the execution subject of step 103 can be device B; and so on.

[0119] In addition, in some of the processes described in the above embodiments and the accompanying drawings, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel, and the sequence numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0120] Figure 3 A structural diagram of a quality assessment model training device provided by an exemplary embodiment of the present application is shown in FIG. Figure 3 As shown, the device includes: an acquisition module 31, a recognition module 32, a construction module 33 and a training module 34.

[0121] An acquisition module 31 is used to acquire at least two sample pages, where the at least two sample pages include a plurality of sample information objects;

[0122] An identification module 32, for identifying, for any sample page, from a plurality of sample information objects, in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page;

[0123] A construction module 33 is used to construct a double contrast loss function based on the samples in the page and the negative samples outside the page, and the double contrast loss function is used for the initial quality assessment model to learn the user's interest in the interaction of the information object and predict the classification boundary of whether the user has an interactive behavior;

[0124] The training module 34 is used to iteratively train the initial quality assessment model according to the dual contrast loss function and the in-page samples and the out-page negative samples to obtain a target quality assessment model.

[0125] In an optional embodiment, the identification module is specifically used to: for any sample page, identify the in-page samples belonging to any sample page and the out-page negative samples not belonging to any sample page from multiple sample information objects based on the page identification information and behavior label information corresponding to the multiple sample information objects; wherein the page identification information is used to identify the sample page to which the sample information object belongs; and the behavior label information is used to identify whether the sample information object has user interaction data.

[0126] Optionally, the identification module is specifically used to: divide the multiple sample information objects into a first sample information object set belonging to any sample page and a second sample information object set not belonging to any sample page according to page identification information corresponding to the multiple sample information objects; use the sample information objects in the first sample information object set as in-page samples belonging to any sample page, the in-page samples including in-page positive samples with positive feedback behavior label information and in-page negative samples with negative feedback behavior label information; use the negative sample information objects with negative feedback behavior label information in the second sample information object set as off-page negative samples belonging to any sample page; wherein the positive feedback behavior label information identifies that the sample information object has user interaction data, and the negative feedback behavior label information identifies that the sample information object does not have user interaction data.

[0127] In an optional embodiment, the construction module is specifically used to: for any sample page, determine the quality score variables within the page corresponding to the sample within the page, and the quality score variables outside the page corresponding to the negative sample outside the page; and construct a double contrast loss function of the initial quality assessment model based on the quality score variables within the page and the quality score variables outside the page of at least two sample pages.

[0128] In an optional embodiment, the training module is specifically used to: for any sample page, use the initial quality assessment model to respectively perform quality assessment on the on-page sample and the off-page negative sample of any sample page to obtain the on-page quality score and the off-page quality score corresponding to any sample page; substitute the on-page quality score and the off-page quality score corresponding to at least two sample pages into the double contrast loss function to obtain the double contrast loss function value;

[0129] When the double contrast loss function value does not satisfy the iteration stop condition, the parameters of the initial quality assessment model continue to be optimized until the iteration stop condition is satisfied to obtain the target quality assessment model.

[0130] Optionally, any sample page also includes sample user information and sample retrieval information; the training module is specifically used to: input the sample user information, sample retrieval information and on-page samples of any sample page into the initial quality assessment model for quality assessment, and obtain the on-page quality score corresponding to any sample page; input the sample user information, sample retrieval information and off-page negative samples of any sample page into the initial quality assessment model for quality assessment, and obtain the off-page quality score corresponding to any sample page.

[0131] Optionally, the quality scores within the page include: a first quality score set adapted to the positive samples within the page included in the samples within the page and a second quality score set adapted to the negative samples within the page included in the samples within the page; the quality scores outside the page include a third quality score set adapted to the negative samples outside the page of any sample page; the training module is specifically used to: for any positive sample within the page corresponding to any sample page, calculate the sub-double contrast loss value of the positive sample within the page according to the first quality score corresponding to any positive sample within the page in the first quality score set, the second quality score set corresponding to any sample page, and the third quality score set; generate the double contrast loss function value corresponding to the initial quality assessment model according to the double comparison sub-loss values ​​of the positive samples within the page included in at least two sample pages.

[0132] Further optionally, the training module is specifically used to: determine the weight information corresponding to the negative samples in the page and the weight information corresponding to the negative samples outside the page according to the second quality score set and the third quality score set corresponding to any sample page; determine the negative sample quality score according to the second quality score set and its corresponding weight information and the third quality score set and its corresponding weight information; generate a sub-double contrast loss value corresponding to the positive sample in any page according to the negative sample quality score and the first quality score corresponding to the positive sample in any page in the first quality score set.

[0133] In another optional embodiment, the training module is specifically used to: obtain a pre-constructed binary classification loss function, which is defined based on multiple quality score variables corresponding to multiple sample information objects; use the double contrast loss function and the binary classification loss function as optimization targets, and use at least in-page samples and out-page negative samples to train the initial quality assessment model to obtain a target quality assessment model.

[0134] Optionally, the training module is specifically used to: use an initial quality assessment model to perform quality scores on multiple sample information objects; substitute the quality scores of the multiple sample information objects into a binary classification loss function to obtain a binary classification loss function value; for any sample page, use the initial quality assessment model to perform quality assessment on the in-page samples and off-page negative samples of any sample page respectively to obtain an in-page quality score and an off-page quality score corresponding to any sample page; substitute the in-page quality scores and off-page quality scores corresponding to at least two sample pages into a double contrast loss function to obtain a double contrast loss function value; when at least one of the binary classification loss function value and the double contrast loss function value does not meet the corresponding iteration stop condition, continue to optimize the parameters of the initial quality assessment model to obtain a target quality assessment model.

[0135] The embodiment of the present application also provides an information search device, which includes: a determination module, an acquisition module, an evaluation module, a selection module and a display module;

[0136] A determination module, configured to determine user information and search information corresponding to a search operation in response to a search operation initiated by a target user;

[0137] An acquisition module, used for acquiring multiple candidate information objects according to the search information;

[0138] An evaluation module, used to evaluate the quality of multiple candidate information objects using a target quality scoring model according to user information and search information, so as to obtain quality scores of each of the multiple candidate information objects;

[0139] A selection module, used to select a target information object from a plurality of candidate information objects according to their respective quality scores; a display module, used to display on a result page corresponding to the search operation;

[0140] Among them, the target quality scoring model is trained based on the above-mentioned quality assessment model training method.

[0141] The embodiment of the present application also provides an information recommendation device, which includes: an acquisition module, an evaluation module, a selection module and a display module;

[0142] An acquisition module, used to acquire multiple candidate information objects when a set event is triggered;

[0143] An evaluation module, used to perform quality evaluation on multiple candidate information objects using a target quality scoring model to obtain quality scores of each of the multiple candidate information objects;

[0144] A selection module is used to select a target information object from a plurality of candidate information objects according to their respective quality scores; a display module is used to display on a target page;

[0145] Among them, the target quality scoring model is trained based on the above-mentioned quality assessment model training method.

[0146] The detailed implementation and beneficial effects of each step in the above-mentioned device provided in the embodiments of the present application have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0147] Figure 4 A schematic diagram of the structure of an electronic device provided by an exemplary embodiment of the present application is shown in FIG. Figure 4 As shown, the device includes: a memory 44 and a processor 45.

[0148] The memory 44 is used to store computer programs and can be configured to store various other data to support operations on the electronic device. Examples of such data include instructions for any application or method used to operate on the electronic device.

[0149] The processor 45 is coupled to the memory 44 and is used to execute the computer program in the memory 44, so as to: obtain at least two sample pages, the at least two sample pages include multiple sample information objects; for any sample page, identify the in-page samples belonging to any sample page and the out-page negative samples not belonging to any sample page from the multiple sample information objects; construct a double contrast loss function based on the in-page samples and the out-page negative samples, the double contrast loss function is used for the initial quality assessment model to learn the user's interest in the interaction of the information object and predict the classification boundary of whether the user has an interactive behavior; according to the double contrast loss function and the in-page samples and the out-page negative samples, the initial quality assessment model is iteratively trained to obtain the target quality assessment model.

[0150] In an optional embodiment, when the processor 45 identifies, for any sample page, from multiple sample information objects, in-page samples belonging to any sample page and out-page negative samples that do not belong to any sample page, the processor 45 is specifically used to: for any sample page, identify, from multiple sample information objects, in-page samples belonging to any sample page and out-page negative samples that do not belong to any sample page, according to page identification information and behavior label information corresponding to the multiple sample information objects; wherein the page identification information is used to identify the sample page to which the sample information object belongs; and the behavior label information is used to identify whether the sample information object has user interaction data.

[0151] Optionally, when the processor 45 identifies in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page from multiple sample information objects based on page identification information and behavior label information corresponding to multiple sample information objects, the processor 45 is specifically used to: divide the multiple sample information objects into a first sample information object set belonging to any sample page and a second sample information object set not belonging to any sample page based on the page identification information corresponding to the multiple sample information objects; use the sample information objects in the first sample information object set as in-page samples belonging to any sample page, the in-page samples including in-page positive samples with positive feedback behavior label information and in-page negative samples with negative feedback behavior label information; use the negative sample information objects with negative feedback behavior label information in the second sample information object set as out-page negative samples belonging to any sample page; wherein the positive feedback behavior label information identifies that the sample information object has user interaction data, and the negative feedback behavior label information identifies that the sample information object does not have user interaction data.

[0152] In an optional embodiment, when the processor 45 constructs a dual contrast loss function based on the in-page sample and the off-page negative sample, it is specifically used to: for any sample page, determine the in-page quality score variable corresponding to the in-page sample, and the off-page quality score variable corresponding to the off-page negative sample; and construct a dual contrast loss function of the initial quality assessment model based on the in-page quality score variable and the off-page quality score variable of at least two sample pages.

[0153] In an optional embodiment, when the processor 45 iteratively trains the initial quality assessment model according to the dual contrast loss function and the in-page samples and the off-page negative samples to obtain the target quality assessment model, it is specifically used to: for any sample page, use the initial quality assessment model to respectively perform quality assessment on the in-page samples and the off-page negative samples of any sample page to obtain the in-page quality score and the off-page quality score corresponding to any sample page; substitute the in-page quality scores and the off-page quality scores corresponding to at least two sample pages into the dual contrast loss function to obtain the value of the dual contrast loss function; when the value of the dual contrast loss function does not meet the iteration stop condition, continue to optimize the parameters of the initial quality assessment model until the iteration stop condition is met to obtain the target quality assessment model.

[0154] In an optional embodiment, any sample page also includes sample user information and sample retrieval information; when the processor 45 uses the initial quality assessment model to perform quality assessment on the in-page samples and off-page negative samples of any sample page respectively to obtain the in-page quality score and the off-page quality score corresponding to any sample page, it is specifically used to: input the sample user information, sample retrieval information and in-page samples of any sample page into the initial quality assessment model for quality assessment to obtain the in-page quality score corresponding to any sample page; input the sample user information, sample retrieval information and off-page negative samples of any sample page into the initial quality assessment model for quality assessment to obtain the off-page quality score corresponding to any sample page.

[0155] Optionally, the quality score within the page includes: a first quality score set adapted to the positive samples within the page included in the samples within the page and a second quality score set adapted to the negative samples within the page included in the samples within the page; the quality score outside the page includes a third quality score set adapted to the negative samples outside the page of any sample page; when the processor 45 substitutes the quality score within the page and the quality score outside the page corresponding to at least two sample pages into the double contrast loss function to obtain the double contrast loss function value, it is specifically used to: for any positive sample within the page corresponding to any sample page, according to the first quality score corresponding to any positive sample within the page in the first quality score set, the second quality score set and the third quality score set corresponding to any sample page, calculate the sub-double contrast loss value of the positive sample within the page; according to the double comparison sub-loss value of the positive sample within the page included in at least two sample pages, generate the double contrast loss function value corresponding to the initial quality assessment model.

[0156] Further optionally, when the processor 45 determines the sub-double contrast loss value of the positive sample in any page according to the first quality score, the second quality score set and the third quality score set corresponding to the positive sample in any page in the first quality score set, the processor 45 is specifically used to: determine the weight information corresponding to the negative sample in the page and the weight information corresponding to the negative sample outside the page according to the second quality score set and the third quality score set corresponding to any sample page; determine the negative sample quality score according to the second quality score set and its corresponding weight information and the third quality score set and its corresponding weight information; generate the sub-double contrast loss value corresponding to the positive sample in any page according to the negative sample quality score and the first quality score corresponding to the positive sample in any page in the first quality score set.

[0157] In an optional embodiment, when the processor 45 iteratively trains the initial quality assessment model according to the dual contrast loss function and the in-page samples and the off-page negative samples to obtain the target quality assessment model, it is specifically used to: obtain a pre-constructed binary classification loss function, which is defined according to multiple quality scoring variables corresponding to multiple sample information objects; use the dual contrast loss function and the binary classification loss function as optimization targets, and at least use the in-page samples and the off-page negative samples to train the initial quality assessment model to obtain the target quality assessment model.

[0158] Optionally, when the processor 45 uses the dual contrast loss function and the binary classification loss function as optimization targets and uses at least the in-page samples and the off-page negative samples to perform model training on the initial quality assessment model to obtain a target quality assessment model, the processor 45 is specifically used to: use the initial quality assessment model to perform quality scores on multiple sample information objects; substitute the quality scores of the multiple sample information objects into the binary classification loss function to obtain the binary classification loss function value; for any sample page, use the initial quality assessment model to perform quality assessment on the in-page samples and the off-page negative samples of any sample page respectively to obtain the in-page quality score and the off-page quality score corresponding to any sample page; substitute the in-page quality scores and the off-page quality scores corresponding to at least two sample pages into the dual contrast loss function to obtain the dual contrast loss function value; when at least one of the binary classification loss function value and the dual contrast loss function value does not meet the corresponding iteration stop condition, continue to optimize the parameters of the initial quality assessment model to obtain the target quality assessment model.

[0159] Regarding the embodiments of the present application Figure 4 The detailed implementation and beneficial effects of each step in the illustrated device have been described in detail in the aforementioned embodiments and will not be elaborated on here.

[0160] Further, if Figure 4 As shown, the electronic device also includes: a communication component 46, a display 47, a power component 48, an audio component 49 and other components. Figure 4 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 4 In addition, Figure 4 The components in the dashed box are optional components, not mandatory components, and the specific components depend on the product form of the electronic device. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, or an IOT (Internet of Things) device, or a server device such as a conventional server, a cloud server, or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it can include Figure 4If the electronic device of this embodiment is implemented as a server device such as a conventional server, a cloud server or a server array, it may not include Figure 4 Components within the dashed box.

[0161] The embodiment of the present application also provides an electronic device, the implementation structure of the electronic device is Figure 4 The implementation structure of the electronic device shown is the same or similar, and can refer to Figure 4 The electronic device provided in this embodiment is implemented as follows: Figure 4 The difference between the electronic devices in the illustrated embodiments mainly lies in: the functions implemented by the processor executing the computer program stored in the memory are different. For the electronic device provided in this embodiment, its processor executes the computer program stored in the memory, which can be used to: determine the user information and search information corresponding to the search operation in response to the search operation initiated by the target user; obtain multiple candidate information objects based on the search information; perform quality assessment on the multiple candidate information objects using the target quality scoring model based on the user information and the search information to obtain the quality scores of the multiple candidate information objects; select the target information object from the multiple candidate information objects based on the quality scores of the multiple candidate information objects, and display it on the result page corresponding to the search operation; wherein the target quality scoring model is based on Figure 1 The method shown is used for training.

[0162] The embodiment of the present application also provides an electronic device, the implementation structure of the electronic device is similar to Figure 4 The implementation structure of the electronic device shown is the same or similar, and can refer to Figure 4 The electronic device provided in this embodiment is implemented as follows: Figure 4 The difference between the electronic devices in the illustrated embodiments mainly lies in that the functions implemented by the processor executing the computer program stored in the memory are different. For the electronic device provided in this embodiment, its processor executes the computer program stored in the memory, which can be used to: obtain multiple candidate information objects when a set event is triggered; use the target quality scoring model to evaluate the quality of the multiple candidate information objects to obtain the quality scores of each of the multiple candidate information objects; select the target information object from the multiple candidate information objects according to the quality scores of each of the multiple candidate information objects, and display it on the target page; wherein the target quality scoring model is based on Figure 1 The method shown is used for training.

[0163] The detailed implementation and beneficial effects of the electronic device provided in the embodiments of the present application have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0164] Accordingly, an embodiment of the present application also provides a computer-readable storage medium storing a computer program, which, when executed, can implement the steps that can be executed by an electronic device in the above-mentioned quality scoring model training method, information search method, and information recommendation method embodiments.

[0165] Accordingly, an embodiment of the present application also provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the processor implements the steps that can be executed by an electronic device in the above-mentioned quality scoring model training method, information search method, and information recommendation method embodiments.

[0166] The above-mentioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0167] The above-mentioned communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wide Band (UWB) technology, Bluetooth (BT) technology and other technologies.

[0168] The above-mentioned display includes a screen, and the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0169] The power supply assembly provides power to various components of the device where the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply assembly is located.

[0170] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (Microphone, MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting an audio signal.

[0171] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-readable storage media (including but not limited to disk storage, compact disc read-only memory (Compact Disc Read-Only Memory, CD-ROM), optical storage, etc.) containing computer-usable program code.

[0172] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0173] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0174] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0175] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interface, network interface and memory.

[0176] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0177] Computer readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0178] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0179] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A quality assessment model training method, characterized in that: include: Acquire at least two sample pages, wherein the at least two sample pages include a plurality of sample information objects; For any sample page, identifying, from the multiple sample information objects, in-page samples belonging to the sample page and out-page negative samples not belonging to the sample page; Constructing a double contrast loss function based on the samples in the page and the negative samples outside the page, wherein the double contrast loss function is used for the initial quality assessment model to learn the user's interest in the interaction with the information object and to predict the classification boundary of whether the user has an interactive behavior; The initial quality assessment model is iteratively trained according to the dual contrast loss function and the in-page samples and the out-of-page negative samples to obtain a target quality assessment model.

2. The method according to claim 1, characterized in that For any sample page, identifying, from the multiple sample information objects, in-page samples belonging to any sample page and out-page negative samples not belonging to any sample page, including: For any sample page, according to the page identification information and behavior label information corresponding to the multiple sample information objects, identify the in-page samples belonging to the any sample page and the out-page negative samples not belonging to the any sample page from the multiple sample information objects; The page identification information is used to identify the sample page to which the sample information object belongs; and the behavior tag information is used to identify whether the sample information object has user interaction data.

3. The method according to claim 2, characterized in that According to the page identification information and the behavior label information corresponding to the multiple sample information objects, identifying the in-page samples belonging to any of the sample pages and the out-page negative samples not belonging to any of the sample pages from the multiple sample information objects, including: According to the page identification information corresponding to the multiple sample information objects, the multiple sample information objects are divided into a first sample information object set belonging to any of the sample pages and a second sample information object set not belonging to any of the sample pages; Using the sample information objects in the first sample information object set as in-page samples belonging to any sample page, the in-page samples including in-page positive samples with positive feedback behavior label information and in-page negative samples with negative feedback behavior label information; taking the negative sample information object with negative feedback behavior label information in the second sample information object set as an off-page negative sample belonging to any of the sample pages; The positive feedback behavior label information identifies that the sample information object has user interaction data, and the negative feedback behavior label information identifies that the sample information object does not have user interaction data.

4. The method according to claim 1, characterized in that: Constructing a dual contrast loss function based on the in-page sample and the out-page negative sample, including: For any sample page, determine the quality score variable on the page corresponding to the sample on the page, and the quality score variable off the page corresponding to the negative sample off the page; A double contrast loss function of the initial quality assessment model is constructed according to the on-page quality score variables and the off-page quality score variables of the at least two sample pages.

5. The method according to claim 1, characterized in that According to the dual contrast loss function and the in-page samples and the out-page negative samples, the initial quality assessment model is trained to obtain a target quality assessment model, including: For any sample page, the initial quality assessment model is used to perform quality assessment on the on-page samples and the off-page negative samples of the sample page, so as to obtain the on-page quality score and the off-page quality score corresponding to the sample page; Substituting the on-page quality scores and the off-page quality scores corresponding to the at least two sample pages into the double contrast loss function to obtain a double contrast loss function value; When the dual contrast loss function value does not satisfy the iteration stop condition, the parameters of the initial quality assessment model continue to be optimized until the iteration stop condition is satisfied, so as to obtain a target quality assessment model.

6. The method according to claim 5, characterized in that Any of the sample pages also includes sample user information and sample search information; Using the initial quality assessment model to perform quality assessment on the on-page samples and the off-page negative samples of any sample page, respectively, to obtain an on-page quality score and an off-page quality score corresponding to any sample page, including: Inputting the sample user information, sample search information and samples in the page of any sample page into the initial quality assessment model for quality assessment to obtain the quality score in the page corresponding to any sample page; The sample user information, sample search information and off-page negative samples of any sample page are input into the initial quality assessment model for quality assessment to obtain an off-page quality score corresponding to any sample page.

7. The method according to claim 5, characterized in that The quality scores within the page include: a first quality score set adapted to the positive samples within the page included in the samples within the page and a second quality score set adapted to the negative samples within the page included in the samples within the page; the quality scores outside the page include a third quality score set adapted to the negative samples outside the page of any sample page; Substituting the on-page quality scores and the off-page quality scores corresponding to the at least two sample pages into the double contrast loss function to obtain the double contrast loss function value includes: For any in-page positive sample corresponding to any sample page, calculating a sub-double contrast loss value of the in-page positive sample according to a first quality score corresponding to the in-page positive sample in the first quality score set, a second quality score set and a third quality score set corresponding to the in-page positive sample in the first quality score set; A double comparison loss function value corresponding to the initial quality assessment model is generated according to the double comparison sub-loss values ​​of the positive samples in the pages included in the at least two sample pages.

8. The method according to claim 7, characterized in that According to the first quality score corresponding to the positive sample in any page in the first quality score set, the second quality score set, and the third quality score set, the sub-double contrast loss value of the positive sample in any page includes: Determining weight information corresponding to the negative sample in the page and weight information corresponding to the negative sample outside the page according to the second quality score set and the third quality score set corresponding to any one of the sample pages; Determining a negative sample quality score according to the second quality score set and its corresponding weight information and the third quality score set and its corresponding weight information; A sub-double contrast loss value corresponding to the positive sample in any page is generated according to the negative sample quality score and the first quality score corresponding to the positive sample in any page in the first quality score set.

9. The method according to claim 1, characterized in that: According to the dual contrast loss function and the in-page samples and the out-page negative samples, the initial quality assessment model is iteratively trained to obtain a target quality assessment model, including: Obtaining a pre-constructed binary classification loss function, where the binary classification loss function is defined according to a plurality of quality score variables corresponding to the plurality of sample information objects; Taking the dual contrast loss function and the binary classification loss function as optimization targets, at least the in-page samples and the out-page negative samples are used to train the initial quality assessment model to obtain a target quality assessment model.

10. The method according to claim 9, characterized in that Taking the dual contrast loss function and the binary classification loss function as optimization targets, at least using the in-page samples and the out-page negative samples to train the initial quality assessment model to obtain a target quality assessment model, including: Using the initial quality assessment model to perform quality scoring on the multiple sample information objects; Substituting the quality scores of the multiple sample information objects into the binary classification loss function to obtain a binary classification loss function value; For any sample page, the initial quality assessment model is used to perform quality assessment on the on-page samples and the off-page negative samples of the sample page, so as to obtain the on-page quality score and the off-page quality score corresponding to the sample page; Substituting the on-page quality scores and the off-page quality scores corresponding to the at least two sample pages into the double contrast loss function to obtain a double contrast loss function value; When at least one of the binary classification loss function value and the dual contrast loss function value does not satisfy the corresponding iteration stop condition, continue to optimize the parameters of the initial quality assessment model to obtain a target quality assessment model.

11. An information search method, characterized in that: include: In response to a search operation initiated by a target user, determining user information and search information corresponding to the search operation; According to the search information, obtaining multiple candidate information objects; According to the user information and the search information, the target quality scoring model is used to perform quality assessment on the multiple candidate information objects to obtain quality scores of the multiple candidate information objects; Selecting a target information object from the multiple candidate information objects according to the quality scores of the multiple candidate information objects, and displaying the target information object on a result page corresponding to the search operation; Wherein, the target quality scoring model is obtained by training based on the method described in any one of claims 1-10.

12. An information recommendation method, characterized in that: include: When a setting event is triggered, multiple candidate information objects are obtained; Using a target quality scoring model to perform quality assessment on the multiple candidate information objects to obtain a quality score for each of the multiple candidate information objects; Selecting a target information object from the plurality of candidate information objects according to the respective quality scores of the plurality of candidate information objects, and displaying the target information object on a target page; Wherein, the target quality scoring model is obtained by training based on the method described in any one of claims 1-10.

13. An electronic device, characterized in that: include: Memory and processor; The memory is used to store a computer program; the processor is coupled to the memory and is used to execute the computer program to implement the steps in the method of any one of claims 1 to 10, claim 11 and claim 12.

14. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to implement the steps of the method according to any one of claims 1 to 10, claim 11 and claim 12.

15. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the processor is caused to implement the steps in the method of any one of claims 1 to 10, claim 11 and claim 12.

Images