Method and device for generating push information

Abstract

The embodiment of the invention provides a method and device for generating push information, on one hand, personalized push information is generated in combination with user information on the basis of seed information, the problems of divergence, illusion and the like of a generation model can be effectively avoided, on the other hand, users are firstly clustered in the seed information generation process, and the generation efficiency is improved. The corresponding seed information is generated according to the user category, and huge workload caused by marking of a large number of users can be avoided. In a word, the implementation mode provided by the specification can improve the effectiveness of personalized information pushing.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the field of computer technology, and in particular, to a method and device for generating push information. Background Art

[0002] The so-called information push is usually a new technology that reduces information overload by actively transmitting information needed by users through certain technical standards or protocols on the Internet through conditional triggers (such as reaching at regular time intervals, opening relevant pages of an application, etc.). The push technology reduces the time spent searching on the network by automatically transmitting information to users and helps users efficiently discover valuable information. Here, push information can be used to describe the candidate information to be pushed to users or the information pushed to users. Push information can concisely and quickly describe the target to be pushed. For example, the advertisement pictures of public service advertisements, event promotion pages, etc. are a type of push information. The advertisement pictures usually include the image information and advertising slogans of the target to be pushed. With the development of artificial intelligence, in order to improve the user experience, more and more information push products pay more attention to personalization, that is, according to the user profile, corresponding candidate information is selected for the user to push.

[0003] Personalized information push can include the following two aspects: selecting one or more from the pre-determined candidate information for different users as the personalized push information to be shown to users; generating personalized push information for users for information push. In the case of generating personalized push information, how to avoid model hallucinations and overly divergent generation results while providing more personalized and targeted push information is a technical problem worthy of research. Summary of the Invention

[0004] One or more embodiments of this specification describe a method and device for generating push information to solve one or more problems mentioned in the background art.

[0005] According to a first aspect, a method for generating push information is provided. The method includes: obtaining user information of a first user; based on the user information of the first user, matching at least one target user category for the first user from each user category, where the user category is determined by clustering multiple users using user information; obtaining each group of seed information corresponding to each target user category, where a single user category corresponds to a single group of seed information; using each group of seed information, the user information of the first user, and business scenario information as inputs to a generation model, and determining the push information to be pushed to the first user according to the output of the generation model.

[0006] In one embodiment, each user category is determined as follows: the user information of the multiple users is clustered in multiple clustering directions, so that several clusters are obtained in a single clustering direction; one cluster is selected from each of the multiple clustering directions for combination to obtain multiple cluster combinations, and a single cluster combination corresponds to a single user category.

[0007] In a further embodiment, the matching of at least one target user category from each user category based on the user information of the first user includes: determining the cluster with the closest distance from the cluster center to the first user in a single clustering direction as a single target cluster; determining the user categories corresponding to the target cluster combinations in each clustering direction as target user categories.

[0008] In a still further embodiment, when there are multiple cluster centers in a single clustering direction whose respective distances from the first user are all less than a first threshold and the difference between the respective distances is less than a second threshold, all of the multiple clusters are determined as the target clusters in that clustering direction; the determining of the user categories corresponding to the target cluster combinations in each clustering direction as target user categories further includes: respectively combining each of the target clusters in that clustering direction with the target clusters in other directions, and determining all of the user categories corresponding to each of the target cluster groups as target user categories.

[0009] In one embodiment, a single piece of seed information includes at least one of text information, image information, and audio information, and a single set of seed information includes at least one piece of seed information, and is generated as follows: inputting the data corresponding to the corresponding single user category into a seed generation model, and generating multiple candidate seed information via the seed generation model, where the data corresponding to a single user category is one of the following: user information described in natural language, a feature vector formed by obtaining the eigenvalue of each cluster center corresponding to the corresponding user category in the corresponding clustering dimension; selecting at least one from each of the candidate seed information as a single set of seed information.

[0010] In one embodiment, the generation model is a large language model, and the inputting of the seed information and the user information of the first user as the input of the generation model includes: converting the user information of the first user into natural language information expressed in natural language; using the seed information, the natural language information, and the business scenario information of the current push service scenario together as the prompt information of the large language model and inputting it into the large language model, where the business scenario information includes at least one of the following: business field, business object, target to be pushed, length and format of the push information.

[0011] In one embodiment, the output of the generation model includes multiple pieces of candidate information. Determining the push information to be pushed to the first user according to the output of the generation model includes: randomly determining one piece of candidate information as the push information to be pushed to the first user; or, using an evaluation model to determine an evaluation score for each piece of candidate information; selecting candidate information in descending order of the evaluation scores as the push information to be pushed to the first user.

[0012] In one embodiment, the method further includes: obtaining operation data of the current user on the pushed information, and using the operation data as a label to fine-tune the generation model when an update condition is met. The update condition includes at least one of the following: the time period arrives, and the quantity of operation data reaches a predetermined quantity.

[0013] According to a second aspect, there is provided an apparatus for generating push information, the apparatus including:

[0014] An obtaining unit configured to obtain user information of a first user;

[0015] A matching unit configured to match at least one target user category for the first user from each user category based on the user information of the first user, where the user categories are determined by clustering multiple users using user information; and,

[0016] Obtaining each group of seed information corresponding to each target user category, where a single user category corresponds to a single group of seed information;

[0017] A generating unit configured to use each group of seed information, the user information of the first user, and service scenario information as inputs to a generation model, and determine push information to be pushed to the first user according to the output of the generation model.

[0018] According to a third aspect, there is provided a computer-readable storage medium, on which a computer program is stored. When the computer program is executed in a computer, the computer is made to execute the method of the first aspect.

[0019] According to a fourth aspect, there is provided a computing device including a memory and a processor. An executable code is stored in the memory, and when the processor executes the executable code, the method of the first aspect is implemented.

[0020] Through the device and method provided by the embodiments of this specification, in the business scenario of information push, multiple users can be clustered according to user information first to obtain multiple user categories (in the case where there is only one clustering direction, a single user category can correspond to a single cluster, and in the case where there are multiple clustering directions, a single user category corresponds to a clustering combination formed by taking one cluster from each clustering direction). And seed information is determined for each user category. In the case of information push, according to the user information, the corresponding target user category can be matched, the seed information corresponding to the target user category is obtained, and together with the user information and business scenario information, it is processed through a generation model (such as a large language model), so as to obtain the push information for pushing information to the current user.

[0021] In this way, performing personalized information prediction based on the seed information is equivalent to having a certain reference or constraint on the prediction result, avoiding problems such as model hallucinations caused by the overly divergent results generated by the generation model, and also reducing the dependence on the capabilities of the generation model. At the same time, in the process of determining the seed information, multiple user categories are obtained through user clustering and grouping, and a single set of seed information is determined for a single user category, which also greatly reduces the workload of determining the seed information. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other accompanying drawings based on these drawings without creative efforts.

[0023] Figure 1 A schematic diagram of a specific implementation architecture under the technical concept of this specification;

[0024] Figure 2 A schematic flowchart showing the process of generating push information according to an embodiment of this specification;

[0025] Figure 3 A schematic diagram of a specific application architecture according to an embodiment of this specification;

[0026] Figure 4 A block diagram showing the structure of the device for generating push information according to an embodiment of this specification. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The following describes the solution provided by this specification with reference to the accompanying drawings.

[0028] Figure 1 A schematic diagram of a specific implementation architecture of this specification is shown. As Figure 1As shown, in the implementation architecture provided in this specification, it includes a business terminal corresponding to the business party, a user terminal corresponding to the user, and a server. Both the business terminal and the user terminal can be various terminal devices that can interact with the user, such as smartphones, laptops, tablets, smartwatches, etc. The business terminal is the terminal used by the business party. In Figure 1 the shown implementation architecture, the business party can be various business entities that push relevant information of the target to be pushed to the user, such as merchants, enterprises and institutions, event organizers, etc. Correspondingly, the target to be pushed can be, for example, products, policies and regulations to be publicized, event invitations, etc. The user terminal can be any terminal device corresponding to the user who may receive the push information. The user terminal can install and run various terminal applications. For example, information push applications, news applications, shopping applications, food delivery platform applications, Q&A platform applications, etc. The server can provide business support for the terminal applications. For example, it can provide business support for news applications, etc. The server can have a computing platform set locally, or be connected to a computing platform on other devices, for determining the business content provided to the user terminal. For example, the server of an information push application can generate personalized push information through the computing platform and push it to the user terminal. For example Figure 1 in, under the same trigger conditions (such as the user opening the application or the business party actively broadcasting), information 1 can be pushed to user 1, and information 2 can be pushed to user 2, and so on.

[0029] It can be understood that in Figure 1 the shown architecture, the server and the business terminal can be different terminals, or they can be the same. That is to say, the business party itself is the service provider that provides services for the user terminal, and the identities of the business party and the user can also be interchanged. For example, in an information push service, the business party acts as the service provider to provide the target to be pushed, and in another information push service, the business party can act as the user to receive the push information. In addition, Figure 1 the shown architecture is only a specific example. In practice, the number of business terminals and the number of user terminals can be any possible numbers, such as 100,000, 100 million, etc., which are not limited here.

[0030] In this specification, in an information push scenario, personalized push information composed of at least one of text, image, and audio is generated.

[0031] In the process of generating personalized push information for users, the push information generated by the generation model combining user information and business scenarios may have a problem of being relatively divergent. Especially when using a large language model as the generation model, the training corpus is relatively rich, and model hallucinations may occur.

[0032] In view of this, this specification provides a technical concept. During the process of personalized generation of user push information, not only user information and scenario information are used as input data for the generation model, but also seed information is provided for the generation model as a reference or constraint, enabling it to be improved based on the seed information according to the scenario information and user information, thereby avoiding the model from being too divergent or having model hallucinations, and making the generated push information more effective.

[0033] The following describes in detail the technical concept of this specification in conjunction with the embodiments shown in the accompanying drawings.

[0034] Figure 2 The flowchart of generating push information according to an embodiment is shown. The execution subject of this process can be a computer, device, or server with certain computing capabilities. More specifically, for example, it is Figure 1 the computing platform shown. When the information push condition is satisfied, personalized push information is generated and pushed to the user through the Figure 2 process shown. Among them, the information push condition can be, for example: the arrival of a predetermined push moment (such as the fixed time point for pushing the daily friend circle exercise step data to the user every day for exercise), the user opening a predetermined page (such as the user opening the browsing page of a certain commodity), the activity initiator initiating a business push application (such as a public welfare organization initiating a push application for a public welfare activity for citizens), and so on.

[0035] As Figure 2 shown, the process of generating push information provided by the embodiments of this specification may include: Step 201, obtaining the user information of the first user; Step 202, based on the user information of the first user, matching at least one target user category for the first user from each user category, where the user categories are determined by clustering multiple users using user information; Step 203, obtaining each group of seed information corresponding to each target user category, where a single user category corresponds to a single group of seed information; Step 204, using each group of seed information, the user information of the first user, and the business scenario information as the input of the generation model, and determining the push information to be pushed to the first user according to the output of the generation model.

[0036] First, in step 201, obtain the user information of the first user.

[0037] Here, the first user can be any network user who meets the information push condition. The network here can be the Internet, or the telecommunications network of a telecommunications operator (such as the GSM network corresponding to a SIM card), which is not limited here. The first user can correspond to a client for the current information push business scenario, and a corresponding client identifier, such as the user ID of an application, the SIM card (card number or IMSI, etc.).

[0038] User information can be various types of information that describe a user, such as one or more of the registration information, behavior information, interest preference information, etc. on a business platform. Among them, the registration information is usually relatively fixed information, such as the user's identity information, phone number, date of birth, and so on. The user's behavior information can include information on various behaviors performed by the user through the device running the corresponding client, such as at least one piece of information among clicks, views, searches, forwards, supports, oppositions, collections, form submissions, locations, etc. obtained from the device operation log. More specifically, the view information can include, for example, information such as the viewed page and the page view duration, and the form submission information can include, for example, shopping information, click information, etc. When the device running the client is a portable mobile device (such as a mobile phone), the user's location information can also correspond to one or more of the movement trajectory information, stay information, stay duration information, and so on. The interest preference information can be information set by the user or information mined through the user's behavior. For example, in a navigation business platform, the user can set the preference information "do not take the highway" by himself, and so on.

[0039] The user information of the first user can be obtained from the client of the first user, or from the business server (such as Figure 1 the server shown, etc.), or part from the client and part from the server, which is not limited here.

[0040] Next, through step 202, at least one target user category is matched for the first user from each user category based on the user information of the first user.

[0041] The user category can be a category determined by clustering multiple users on the business platform using user information. Under the technical concept of this specification, it is necessary to determine the seed information according to the user category. Therefore, it is necessary to determine the user category for the first user. The user category to which the first user belongs is denoted as the target user category. The target user category is at least one. The determination method of the target user category is related to the clustering method of the user category.

[0042] According to some possible implementation methods, using user information, multiple portrait feature dimensions can be artificially determined, corresponding feature values can be extracted from each portrait feature dimension, and clustering can be performed based on the feature vectors composed of these feature values. For example, the multiple portrait dimensions include at least one of the following: occupation, date of birth, consumption preference, average monthly consumption amount, browsing preference (type of viewed page), click preference (type of clicked hyperlink), form submission preference (such as shopping type, direct submission, centralized submission, etc.), and so on. A single user can correspond to a feature vector, and the clustering between feature vectors can be performed through vector similarity (measured by metrics such as cosine similarity, Euclidean distance, etc.).

[0043] Vector clustering methods include, for example, K-means, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), Fuzzy Clustering, Mean Shift Clustering, Hierarchical Clustering (a clustering algorithm based on a tree structure), and so on.

[0044] Taking K-means clustering as an example, K clustering centers can be randomly specified among users. Calculate the vector similarity between the feature vectors of other users and the feature vectors of the K clustering centers, and divide each user into the cluster corresponding to the clustering center with the highest feature vector similarity. Then, re-determine the clustering centers for each current cluster, and repeat the above clustering process until the clustering centers of each cluster converge. Here, the convergence of the clustering centers can be understood as that the change distance of the clustering centers does not exceed a predetermined distance threshold, or the feature vector similarity between the new clustering center and the clustering center of the previous round exceeds a predetermined similarity threshold. In this way, multiple clusters are obtained, and each cluster serves as each user category respectively. Each clustering center can be used as the center point of the corresponding user category in the feature space, serving as the basis for matching user categories.

[0045] At this time, the method for determining user categories is to cluster users according to feature vectors, and the data referred to for clustering can be feature vectors. When matching the corresponding user category for the first user, the feature vector of the first user can be extracted from the user information of the first user, and the distance between the first user and each clustering center (corresponding to a feature vector) can be calculated using the feature vector. In one example, the clustering center closest to the first user can be determined as the corresponding user category. In another example, considering that the first user may be close to the distances of multiple clustering centers, when the distances between each of the multiple clustering centers and the first user are all less than a first threshold (such as 0.1), and the difference between each distance is less than a second threshold (such as 0.01), all the user categories corresponding to the corresponding multiple clusters can be determined as the target user categories.

[0046] According to some other possible implementations, user information can be used for clustering in multiple clustering dimensions (or clustering directions), obtaining multiple clusters in a single clustering dimension, and the cluster categories in each clustering dimension are combined arbitrarily. Each combination corresponds to a single user category. As an example, for instance, each portrait dimension in the foregoing text, such as occupation, date of birth, consumption preference, average monthly consumption amount, browsing preference, click preference, form submission preference, etc., can be used as each clustering dimension respectively. In this way, the user category space can be expanded and the granularity of user category division can be refined. For example, if there are 3 clustering dimensions and the number of clusters obtained are m, n, and k respectively, the size of the user category space can be m×n×k. In an alternative embodiment, for user information for which it is difficult to clearly determine the clustering direction, it can be semantically embedded to obtain a corresponding embedding vector, and the embedding vector can be used as a clustering dimension for clustering to obtain corresponding multiple cluster centers.

[0047] At this time, the way to match the target category for the first user can be: determining the cluster with the closest distance between the cluster center and the first user in a single clustering direction as a single target cluster, and the user category corresponding to the combination of each target cluster can be the target user category. For example, if there are three clustering dimensions, the target cluster a1 is matched in the first clustering dimension, the target cluster b1 is matched in the second clustering dimension, and the target cluster c1 is matched in the third clustering dimension, then the user category corresponding to the combination (a1, b1, c1) of each target cluster can be used as the target user category.

[0048] In one embodiment, considering that the first user may be close to multiple cluster centers in a single clustering dimension, in the case where there are multiple cluster centers corresponding to the first user respectively in a single clustering direction and each distance is less than a first threshold, and the difference between each distance is less than a second threshold, the corresponding multiple clusters are determined as the target clusters in this clustering dimension, and each target cluster in this clustering direction is combined with the target clusters in other directions respectively to form each target cluster group. Thus, the user category corresponding to each target cluster group can be determined as the target user category.

[0049] Under other user category determination methods, the target user category corresponding to the first user can also be determined through corresponding reasonable methods, which will not be listed one by one here.

[0050] It should be noted that the distance calculation between any two vectors involved above can all be performed in a manner inversely proportional to the vector similarity or directly proportional to the vector distance. Among them, the vector similarity is, for example, one of cosine similarity, dot product similarity, etc., and the vector distance is, for example, one of Euclidean distance, variance, Hamming distance, and so on.

[0051] Then, according to step 203, each group of seed information corresponding to each target user category is obtained.

[0052] It can be understood that a single user category can correspond to a single set of seed information. The number of seed information in a single set of seed information is one or more, and a single piece of seed information includes at least one of text, image, and audio information. The seed information can be determined in advance. The determination method of the seed information can be determined manually or generated by a seed generation model (such as a large language model). Taking the generation of seed information by a seed generation model as an example, the eigenvalue of the clustering center can be used as the feature of the corresponding user category and input into the seed generation model, so as to determine the seed information according to the output result of the seed generation model. The number of a set of seed information can be one or multiple.

[0053] According to one embodiment: in the case of clustering according to the feature vector during the process of determining the user category, a single clustering category corresponds to a single user category, and the feature of the clustering center is described by the feature vector, then the feature vector of the clustering center can be used as the feature of the corresponding user category and input into the generation model; in the case of clustering according to the clustering dimension during the process of determining the user category, a single user category corresponds to the feature combination of the clustering centers on each clustering dimension, then the corresponding feature combination of the clustering centers corresponding to a single user category on each clustering dimension can be used as the feature of the corresponding user category and input into the seed generation model, and the candidate seed information is generated by the seed generation model. For example, the clustering dimension includes three dimensions a, b, and c. The clustering center a of dimension a 1 has a single eigenvalue a 11 , the clustering center b of dimension b 1 has a single eigenvalue b 12 , the clustering center c of dimension c 1 has two eigenvalues c 11 、c 12 , then the user category features corresponding to the clustering centers a 1 、b 1 、c 1 can have four dimensions, such as (a 11 , b 12 , c 11 , c 12 ).

[0054] It can be understood that in order to combine the seed information with the business scenario, in some possible designs, the input of the seed generation model can also include the scenario features of the current business scenario. The scenario features of the current business scenario can be extracted based on the description information of the current business scenario. The description information of the current business scenario can be at least one piece of information such as the description of the business domain, business object, target to be pushed, length and format of the push information, etc. The business domain is, for example, product push, news push, event (such as public welfare activities, etc.) push, etc. The business object is, for example, the object targeted by the information push (such as office workers), the effect expected by the business (such as 40% of users can accept, etc.), etc. The target to be pushed is, for example, products, news, events, etc. The length and format of the push information are requirements such as the number of words, whether it includes images, audio, etc.

[0055] The extraction of business scenario features can be carried out by various extraction methods in conventional technologies, which will not be elaborated here. In an optional embodiment, the scenario description information can be input into the embedding layer of the large language model, and the embedding layer of the large language model processes it to obtain the scenario features.

[0056] The scenario features can be shared by multiple users, and under the same business scenario, under the triggering conditions of each user at different times, there is no need to update. Therefore, the scenario features can be determined in the current process, or can be determined in advance and applied to each user until they are updated. The scenario features and user category features can be used as an input of the seed generation model respectively, or can be fused through methods such as weighting, summing, splicing, etc. and then used as the input of the seed generation model, which is not limited here.

[0057] According to another embodiment, when the seed generation model is a natural language processing model, such as BERT, large prediction model, etc., the business scenario information and the user information of each clustering center corresponding to a single user category can also be input into the seed generation model in the form of information described in natural language, and the seed generation model outputs candidate seed information, which will not be elaborated here.

[0058] The candidate seed information generated by the seed generation model can be one or more. In the case of one candidate seed information, it can be directly used as the corresponding seed information. In the case of multiple candidate seed information, all candidate seed information, or a predetermined number (such as 1) of candidate seed information can be selected as seed information by manual, evaluation model or large model. Among them, in the case of preferential selection by a large model, the format, sentence fluency, scene inclusion degree and other contents of the candidate seed information can be detected by the large model, so as to select the better candidate seed information as the seed information. The evaluation model can be a pre-trained scoring model. The data used as the input of the evaluation model in the training sample can include, for example, user features, scene features and candidate seed information, and the output is the evaluation score of each candidate seed information. The evaluation score of the sample is supervised by the manually marked score label, so as to adjust the pending parameters to obtain the evaluation model. When selecting seed information through an evaluation model, user type characteristics, scene characteristics and candidate seed information can be used together as input data of the evaluation model, and the evaluation model outputs an evaluation score for each candidate seed information, thereby selecting a predetermined number (such as 1) of candidate seed information whose evaluation score is greater than a preset threshold or ranked top from large to small as seed information.

[0059] As can be seen from the above description, a single user category corresponds to one or more pieces of seed information, which can be referred to as a set of seed information. Each user category and each corresponding set of seed information are stored in correspondence. For one or more target user categories determined by the first user, a single target user category can correspond to a single set of seed information, so that each set of seed information corresponding to each target user category can be obtained for the first user.

[0060] Furthermore, in step 204, each group of seed information, the user information of the first user, and the business scenario information are used as inputs of the generation model, and the push information to be pushed to the first user is determined according to the output of the generation model.

[0061] It can be understood that the generation model here can be the same or different from the seed generation model mentioned above. The generation model can be implemented by various models that can achieve the generation goal, such as the Bert model for generating text, the adversarial generative network GAN for generating images, and the large language model. The generation model here needs to be improved on the basis of the seed information to generate push information suitable for the first user. The input data used is the user information of the first user, the scenario information of the current business scenario, and the seed information of the user category to which the first user belongs.

[0062] Among them, the user information of the first user, the scenario information of the current business scenario, and the seed information of the user category to which the first user belongs can be input into the generation model in the form of an embedded vector, or can be input into the generation model in the form of a natural language (such as when the generation model is a large language model). In the case of inputting the generation model in the form of an embedded vector, each embedded vector can be fused by splicing, weighting, etc. and then input into the generation model. In the case of inputting the generation model in the form of a natural language, the user information, scenario information, and seed information of the first user can all be converted into a natural language description. The copy is usually in the form of a natural language. When the seed information contains images and audio, the images and audio can be converted into natural language. The user information and scenario information of the first user may contain data formats such as data tables and key-value pairs, which can be converted into natural language forms manually or by a large model.

[0063] The candidate information generated by the generation model may be information including at least one of text, image, and audio, such as text information, graphic promotional copy, video, animation, etc. Since the large language model has good natural language processing capabilities, it can be used as a generation model for text and images, that is, the large language model can be used as a generation module to process the user information, seed information, and scenario information of the current business scenario of the first user, thereby generating candidate information for information push to the first user.

[0064] The generation model can output one candidate information or multiple candidate information. Where one candidate information is output, the candidate information can be directly used as the push information to be pushed to the first user. Where multiple candidate information is output, it can be information generated once or multiple times. At this time, a predetermined number (such as 1) of candidate information can be selected from the multiple candidate information as the push information to be pushed to the first user. The selection method can be: random selection, thereby providing more possibilities for collecting user preferences; preferential selection, such as determining the evaluation score through an evaluation model, selecting the candidate information whose evaluation score exceeds the predetermined score threshold, or ranking the candidate information from large to small; manual designation; and the like.

[0065] According to some possible designs, before determining the candidate information as the push information to be pushed to the above-mentioned first user, quality inspection can be performed first to screen out candidate information with poor quality. The content of the quality inspection includes, for example, at least one of whether the format meets the requirements, the smoothness of the copywriting, the degree of inclusion of scene features, etc. Among them, the degree of scene inclusion can also be understood as the degree of association with the current business scene, which can be determined by the matching degree between the candidate information and the description information of the current business scene. For example, as an example, the candidate information and the description information of the current business scene can be semantically embedded to obtain semantic vectors, and the similarity of the semantic vectors can be calculated. The higher the similarity of the semantic vectors, the higher the degree of inclusion of scene features.

[0066] According to a possible design, during the subsequent information push process, based on the data of the operations (such as click, browse, submit form, delete, etc.) performed by the user on the push information, it can be used as a label for whether the generated push information is effective, and the generation model can be fine-tuned with supervision (SFT) according to this label. In the case of using an evaluation model, the user operation data can also be converted into an evaluation score to jointly fine-tune the generation model and the evaluation model.

[0067] To further clarify the technical concept of this specification, based on the above description, refer to Figure 3 a specific application example of Figure 3 The steps in the dotted box in

[0068] are auxiliary operations outside the information push process, which can be performed in advance (such as preprocessing steps) or performed according to preset conditions after multiple information push processes (such as the model SFT step).

[0069] In Figure 3In the example, for a specific business scenario, data preprocessing is first performed. The preprocessing process includes a user clustering step and a seed information tagging step. Through the user clustering step, multiple users are clustered to obtain multiple user categories; through the seed information tagging step, tagging can be performed in combination with the description information of the current business scenario, so as to respectively determine corresponding groups of seed information for each user category. For example, the user categories and the seed information can be stored in a one-to-one correspondence.

[0070] Continuing to refer to Figure 3 As shown, when it is necessary to push information to a specific user, the user information can be first obtained, and category matching can be performed according to the user information, that is, the corresponding user category is matched for the current user in each user category, so as to utilize the seed information corresponding to the matched user category obtained in the preprocessing process. On the other hand, natural language processing can also be performed on the user information, that is, the user information is converted into a form described in natural language. Then, the seed information, the business scenario description information, the seed information, and the natural language processing result of the user information are used as prompt information and input into a large language model serving as a generation model, and the large language model outputs multiple candidate information. Next, information selection is performed, and quality inspection is carried out manually or by a large model, or scoring is carried out via an evaluation model, so as to screen out the push information suitable for the current user from the multiple candidate information and push the information to the user.

[0071] According to a possible design, after the information is pushed to the user, user operations such as clicks, browsing, and form submissions can be collected, and based on the user operations, supervised fine-tuning (i.e., model SFT) can be provided for the large language model.

[0072] It can be understood that Figure 3 is a specific example of the embodiment described in Figure 2, Figure 2 and various descriptions of the technical concept of this specification can be applied to Figure 3 the example in

[0073] Reviewing the above process, for the method of generating push information provided under the technical concept of this specification, on the one hand, personalized push information is generated by combining user information on the basis of seed information, which can effectively avoid problems such as divergence and hallucination of the generation model, and improve the accuracy of the generated push information in combination with the business scenario and the user. On the other hand, in the process of generating seed information, users are first clustered, and corresponding seed information is generated according to user categories, which can avoid the huge workload caused by tagging a large number of users. In short, the technical concept provided in this specification determines seed information to provide reference and constraints for the push information generated for users on the basis of increasing as little workload as possible, so that the generated push information accurately fits the user characteristics and business scenario, and improves the effectiveness of information push.

[0074] According to an embodiment of another aspect, a device for generating push information is further provided. The device may be disposed in a computer, a terminal, or a server having a certain computing power. More specifically, it is, for example, Figure 1 the computing platform shown. Figure 4 FIG. Figure 4 shows a device 400 for generating push information according to an embodiment. As Figure 4 shown, the device 400 may include:

[0075] An obtaining unit 401, configured to obtain user information of a first user;

[0076] A matching unit 402, configured to match at least one target user category for the first user from each user category based on the user information of the first user, where the user categories are determined by clustering multiple users using user information; and obtain each group of seed information corresponding to each target user category, where a single user category corresponds to a single group of seed information;

[0077] A generating unit 403, configured to use each group of seed information, the user information of the first user, and business scenario information as inputs to a generating model, and determine push information to be pushed to the first user according to the output of the generating model.

[0078] In one embodiment, the device 400 further includes a classification unit (not shown), configured to determine each user category in the following manner: cluster the user information of multiple users in multiple clustering directions, so as to obtain several clusters in a single clustering direction; combine one cluster taken from each of the multiple clustering directions to obtain multiple cluster combinations, and a single cluster combination corresponds to a single user category.

[0079] In a further embodiment, the matching unit 402 may further be configured to: determine the cluster with the closest distance to the first user among the cluster centers in a single clustering direction as a single target cluster; determine the user category corresponding to the target cluster combination in each clustering direction as the target user category.

[0080] In a still further embodiment, when, in a single clustering direction, the distances between multiple cluster centers and the first user respectively are all less than a first threshold, and the difference between the distances is less than a second threshold, the matching unit 402 may further be configured to: determine multiple clusters as the target clusters in this clustering direction; combine each of the target clusters in this clustering direction with the target clusters in other directions, and determine each user category corresponding to each target cluster group as the target user category.

[0081] In one embodiment, the single seed information includes at least one of text, image, and audio. A single set of seed information includes at least one piece of seed information. The apparatus 400 further includes a copywriting generation unit (not shown), configured to generate a single set of seed information in the following manner: input the data corresponding to the respective single user category into a seed generation model, generate multiple candidate seed information via the seed generation model, and the data corresponding to the single user category is one of the following: user information described in natural language, a feature vector formed by obtaining the eigenvalue of each cluster center corresponding to the respective user category on the corresponding clustering dimension; select at least one from each candidate seed information as the single set of seed information.

[0082] In one embodiment, the generation model is a large language model, and the generation unit 403 is further configured to: convert the user information of the first user into natural language information expressed in natural language; use the seed information, natural language information, and the business scenario information of the current push service scenario together as the prompt information of the large language model and input it into the large language model. The business scenario information includes at least one of the following: business field, business object, target to be pushed, length and format of the push information.

[0083] In one embodiment, the output of the generation model includes multiple candidate information, and the generation unit 403 can also be configured to: randomly determine one candidate information as the push information to be pushed to the first user; or, use an evaluation model to determine an evaluation score for each candidate information; select candidate information in descending order of the evaluation score as the push information to be pushed to the first user.

[0084] In one embodiment, the apparatus 400 further includes a fine-tuning unit, configured to: obtain the operation data of the current user on the pushed information, and use the operation data as a label to fine-tune the generation model when the update condition is met. The update condition may include at least one of the following: the time period arrives, the number of operation data reaches a predetermined number, and so on.

[0085] It should be noted that Figure 4 the illustrated apparatus 400 corresponds to Figure 2 the described method, Figure 2 and the corresponding descriptions in the illustrated method embodiments also apply to the apparatus 400 and will not be elaborated here.

[0086] According to an embodiment of another aspect, there is also provided a computer-readable storage medium, on which a computer program is stored. When the above computer program is executed in a computer, the computer is made to execute the method described in conjunction with Figure 2 etc.

[0087] According to an embodiment of still another aspect, a computing device is further provided, including a memory and a processor. An executable code is stored in the memory. When the processor executes the above executable code, the method described in combination with Figure 2 etc. is implemented. Those skilled in the art should be able to realize that in the above one or more examples, the functions described in the embodiments of this specification can be implemented by hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0088] The specific implementation manners described above further elaborate on the purpose, technical solutions, and beneficial effects of the technical concept of this specification. It should be understood that the above description is only the specific implementation manners of the technical concept of this specification and is not used to limit the protection scope of the technical concept of this specification. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions in the embodiments of this specification should be included in the protection scope of the technical concept of this specification.

Claims

1. A method for generating push information, the method comprising: Obtain user information of the first user; matching at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user category is determined by clustering a plurality of users using user information; Obtaining each set of seed information corresponding to each target user category, wherein a single user category corresponds to a single set of seed information; Each group of seed information, the user information of the first user, and the business scenario information are used as inputs of the generation model, and the push information to be pushed to the first user is determined according to the output of the generation model.

2. The method of claim 1, wherein: The individual user categories are determined as follows: Clustering the user information of the multiple users in multiple clustering directions, thereby obtaining multiple clusters in a single clustering direction; One cluster is taken from each of the multiple clustering directions for combination to obtain multiple cluster combinations, and a single cluster combination corresponds to a single user category.

3. The method of claim 2, wherein: The matching at least one target user category from various user categories for the first user based on the user information of the first user includes: Determine a cluster whose cluster center is closest to the first user in a single clustering direction as a single target cluster; The user category corresponding to the target cluster combination in each cluster direction is determined as the target user category.

4. The method of claim 3, wherein: In a case where the distances between multiple cluster centers and the first user in a single cluster direction are all smaller than a first threshold, and the difference between the distances is smaller than a second threshold, the multiple clusters are all determined as target clusters in the cluster direction; The step of determining the user category corresponding to the target cluster combination in each cluster direction as the target user category further includes: Each target cluster in the clustering direction is combined with the target clusters in other directions respectively, and each user category corresponding to each target cluster group is determined as the target user category.

5. The method of claim 1, wherein: A single piece of seed information includes at least one of text, image, and audio. A single set of seed information includes at least one piece of seed information and is generated in the following manner: Inputting data corresponding to a corresponding single user category into a seed generation model, and generating multiple candidate seed information through the seed generation model, wherein the data corresponding to a single user category is one of the following: user information described in natural language, and a feature vector consisting of feature values ​​of each cluster center corresponding to the corresponding user category on a corresponding cluster dimension; At least one is selected from each candidate seed information as a single set of seed information.

6. The method of claim 1, wherein: The generation model is a large language model, and taking the seed information and the user information of the first user as inputs of the generation model includes: converting the user information of the first user into natural language information expressed in a natural language; The seed information, the natural language information, and the business scenario information of the current pushed business scenario are used together as prompt information of the large language model to input the large language model, and the business scenario information includes at least one of the following: business field, business object, target to be pushed, length and format of the pushed information.

7. The method of claim 1, wherein: The output of the generation model includes multiple candidate information, and the push information to be pushed to the first user is determined according to the output of the generation model, including: randomly determine a candidate information as the push information to be pushed to the first user; or Using the evaluation model to determine the evaluation score for each piece of candidate information; The candidate information is selected according to the order of the evaluation scores from high to low as the push information to be pushed to the first user.

8. The method of claim 1, wherein: The method further comprises: The operation data of the current user on the pushed information is obtained, and is used to fine-tune the generation model using the operation data as a label when an update condition is met. The update condition includes at least one of the following: the time period has arrived, and the amount of operation data has reached a predetermined amount.

9. A device for generating push information, the device comprising: an acquisition unit, configured to acquire user information of a first user; a matching unit configured to match at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user category is determined by clustering a plurality of users using user information; as well as, Obtaining each set of seed information corresponding to each target user category, wherein a single user category corresponds to a single set of seed information; The generation unit is configured to use each group of seed information, the user information of the first user, and the business scenario information as inputs of the generation model, and determine the push information to be pushed to the first user according to the output of the generation model.

10. A computing device comprising a memory and a processor, characterized in that: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 8 is implemented.

Images