Message pushing method, medium, computer equipment and program product

Abstract

A message pushing method, a medium, a computer device and a program product, the method comprising: acquiring a pre-generated label set, the label set comprising a plurality of global labels, the global labels being associated with user information and commodity information respectively, the global tag is used for describing association information between a user of the e-commerce platform and a commodity of the e-commerce platform; determining a target global tag from the tag set, and obtaining target commodity information associated with the target global tag; generating a to-be-pushed message which corresponds to the target global tag and comprises the target commodity information through a large language model; and pushing the to-be-pushed message to a target user, wherein the user information of the target user is matched with the user information associated with the target global tag.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a message push method, medium, computer device, and program product. Background Art

[0002] In push messaging (PUSH) scenarios, e-commerce platforms often push various messages to users to attract their attention and encourage purchases. In related technologies, a common practice is to pre-set several themes, such as common promotional themes and "Double 11" shopping spree themes. For each pre-set theme, a relatively fixed and single copywriting template is generated. The specific product name is then simply filled in the corresponding position of the template to generate the push message. However, this traditional and mechanical message generation method has obvious flaws, and the generated messages lack sufficient diversity. Summary of the Invention

[0003] In a first aspect, an embodiment of the present application provides a message push method, the method comprising:

[0004] Obtain a pre-generated tag set, wherein the tag set includes a plurality of global tags, wherein the global tags are respectively associated with user information and product information, and the global tags are used to describe association information between users of the e-commerce platform and products of the e-commerce platform;

[0005] Determine a target global tag from the tag set, and obtain target product information associated with the target global tag;

[0006] Generate a message to be pushed corresponding to the target global tag and including the target product information through a large language model;

[0007] The message to be pushed is pushed to a target user, wherein the user information of the target user matches the user information associated with the target global tag.

[0008] In a second aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present application.

[0009] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any embodiment of the present application when executing the computer program.

[0010] In a fourth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any embodiment of the present application when executed by a processor.

[0011] In an embodiment of the present application, a message to be pushed is generated based on a global tag. Unlike product information that mainly focuses on the characteristics of the product itself and user information that mainly focuses on the characteristics of the user itself, a global tag is used to describe the association information between the users of the e-commerce platform and the products of the e-commerce platform. The information described is no longer limited to the characteristics of the product itself and the user itself, but can reflect the characteristics of the product from the perspective of different user preferences and the position and role of the product in the user behavior chain, and generate new and meaningful tag combinations. Therefore, the fused global tag effectively expands the tag content compared to the original user information and product information, and is more diverse, so that the messages to be pushed generated based on the global tag are also more diverse.

[0012] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings herein are incorporated into the specification and constitute a part of this application. These drawings illustrate embodiments consistent with this application and, together with the specification, are used to illustrate the technical solutions of this application.

[0014] Figure 1 It is a schematic diagram of a message push scenario in an embodiment of the present application.

[0015] Figure 2 This is a flowchart of the message push method of an embodiment of the present application.

[0016] Figure 3A and Figure 3B They are schematic diagrams of the process of obtaining global tags in embodiments of the present application.

[0017] Figure 4A It is an overall flow chart of an embodiment of the present application.

[0018] Figure 4B This is an overall flow chart of the process of obtaining global tags in an embodiment of the present application.

[0019] Figure 4C This is an overall flow chart of the process of obtaining global tags according to another embodiment of the present application.

[0020] Figure 5 It is a schematic diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0022] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of.

[0023] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0024] In order to enable people in this technical field to better understand the technical solutions in the embodiments of the present application, and to make the above-mentioned purposes, features and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are further described in detail below with reference to the accompanying drawings.

[0025] 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 relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0026] In the message push (PUSH) scenario, e-commerce platforms often push various messages to users to attract their attention and purchases. Figure 1 Schematic diagrams of message push scenarios in some embodiments are shown. Figure 1As shown, the message push system may include a server 102 and a client 104, wherein the server 102 is mainly responsible for the generation, screening, scheduling and initiation and execution control of message push tasks. The specific functions of the server 102 are as follows:

[0027] Message Generation: When an e-commerce platform experiences an event such as a product promotion (e.g., a limited-time discount, a discount for purchases above a certain amount), a change in order status (pending payment, shipping, receipt, etc.), a user care reminder (birthday benefits, member-only events), or a new product recommendation, server 102 can automatically generate corresponding message content based on pre-set logical rules and templates. For example, when a popular electronic product launches a limited-time flash sale, server 102 will combine and generate a message text containing key elements of the event to attract users to click and buy, based on the product's detailed information (name, specifications, price, inventory, etc.), the event duration, the discount strength, and other parameters, such as "[e-commerce platform name] [product name] launches a limited-time special offer, with a direct discount of [X] yuan. Only [specific duration] left. Click to view details [link]."

[0028] User screening: Based on multiple factors, such as user profile data, historical behavior records (browsed product categories, purchase frequency, favorites, and regional information), membership level, and spending amount, we accurately screen target user groups that meet specific message push criteria. For example, if a new product is a high-end skincare set, server 102 will prioritize users who have a history of beauty and skincare product purchases, follow the brand's dynamics, and have high spending power to improve the accuracy and conversion rate of message push.

[0029] Message Scheduling and Push: Message push tasks are scheduled based on factors such as message urgency, priority (e.g., system notifications take precedence over regular promotional messages), and target users' active hours (determined by analyzing past login and browsing patterns). During appropriate time windows, server 102 pushes messages in batches or in real time to the receiving interface specified by client 104 via an established secure connection channel (e.g., HTTPS, ensuring secure data transmission) using appropriate push technologies (e.g., persistent or short connections) and protocols (e.g., WebSocket, MQTT). The server also records the execution status of the push task (success, failure, partial delivery, etc.) for subsequent statistical analysis and retry processing.

[0030] The client 104 is mainly responsible for receiving, displaying and interactively responding to messages. Its specific functions are as follows:

[0031] Message Receiving: When the client 104 application (including mobile apps, web versions, and other terminal types) is started or running in the background, it will open a dedicated message receiving port or listener to maintain a real-time communication connection with the server 102. When the server 102 pushes a message, the client 104 can parse and verify the received message data based on a predefined communication protocol to ensure the integrity and accuracy of the message. If the message passes verification, it will be temporarily stored in the local message cache queue, awaiting subsequent display and processing. If the message is found to be damaged, incomplete, or otherwise abnormal, a retry request will be sent to the server 102 to request that the message be pushed again.

[0032] Message display: Based on the type of message (notification, promotion, interaction, etc.), priority, and application scenario, the message will be presented to the user in an appropriate display format at an appropriate location on the client 104. For example, system notification messages will pop up in the form of eye-catching pop-up windows on the application homepage or designated function page; promotional messages will be displayed in the form of push cards in the product recommendation area and the message center page of the personal center, containing eye-catching pictures, concise and clear text descriptions, and corresponding action buttons (such as "view details" and "buy now"). For some important and urgent messages (such as order status change reminders), sound reminders or vibration reminders may also be provided to ensure that users notice them in time. At the same time, the client 104 will differentiate the display style of messages according to the user's reading status (read, unread). For example, unread messages can be highlighted in bold or with an unread mark, making it easier for users to quickly understand important information.

[0033] Interactive Response: When a user clicks to view, close, bookmark, or forward a pushed message, client 104 captures the user's interactive behavior, records relevant operational data (such as click time, click device information, and message content-related operations), and feeds this data back to server 102 through real-time communication with server 102 for subsequent data statistical analysis and processing. For example, when a user clicks the "Buy Now" button in a promotional message, the client will jump to the corresponding product details page and send the user's click behavior data to server 102. Server 102 will then update the user's purchase intention record, activity participation data, and other data based on this data, providing data support for subsequent precision marketing, product recommendations, and other services.

[0034] In related technologies, the common practice for generating push messages is to pre-set several themes, such as common promotional themes and "Double 11" shopping spree themes. For each pre-set theme, a relatively fixed and single copywriting template is generated. Then, the specific product name is simply filled in the corresponding position of the template to generate the push message. However, this traditional and mechanical message generation method has obvious drawbacks:

[0035] From a content perspective, due to the relatively fixed copy templates, where only the product names are changed, the messages are extremely similar in terms of expression, style, and information richness, making them difficult to bring freshness and appeal to users. After repeatedly receiving such stereotyped messages, users are prone to aesthetic fatigue, which can lead to a decrease in their attention to the platform's push notifications, and they may even ignore or choose to turn off push notifications.

[0036] From a user experience perspective, this single push notification method struggles to meet the personalized needs of different users. Different users have varying interests, hobbies, spending habits, and shopping scenarios. However, messages generated using fixed templates often fail to accurately align with users' individual preferences and effectively stimulate their purchasing desire and engagement.

[0037] In summary, the message generation method of traditional e-commerce platforms in the message push scenario, which is based on fixed theme templates and simple product name filling, lacks sufficient diversity.

[0038] Based on this, the embodiment of the present application proposes a message push method, such as Figure 2 As shown, the method includes:

[0039] Step S12: Obtain a pre-generated tag set, which includes multiple global tags. The global tags are associated with user information and product information respectively. The global tags are used to describe the association information between users of the e-commerce platform and products on the e-commerce platform;

[0040] Step S14: determining a target global tag from the tag set, and obtaining target product information associated with the target global tag;

[0041] Step S16: Generate a message to be pushed corresponding to the target global tag and including target product information through the large language model;

[0042] Step S18: Push the message to be pushed to the target user, where the user information of the target user matches the user information associated with the target global tag.

[0043] The embodiment of the present application obtains a global tag for describing the association information between users of the e-commerce platform and the products of the e-commerce platform, and generates a message to be pushed based on the global tag. The original product information mainly focuses on the various characteristics of the product itself, such as brand, price range, specification parameters, functions, categories, etc. This information is relatively direct and specific, and mainly serves the purpose of classifying and describing products. It plays an important role in scenarios such as product management and basic search recommendations, but the dimensions it covers are relatively limited, and mainly revolve around the product itself. When integrated with user information, the situation changes significantly:

[0044] Introducing the user preference dimension: User information includes interests, hobbies, consumption preferences, and purchasing habits. After integration, global tags are no longer limited to the characteristics of the product itself; they can also reflect the characteristics of the product from the perspective of different user preferences. For example, a sportswear product with a "fashionable sporty style" can be associated with a more detailed and biased tag such as "sportswear styles preferred by young male users who enjoy outdoor adventures" by integrating with user information.

[0045] Expanding behavioral association dimensions: By integrating behavioral data from user profiles, such as browsing history and purchase frequency, into product information, global tags can now reflect the product's position and role in the user's behavioral chain. For example, after integration, an electronic product could be labeled as "a product frequently viewed as a photo aid for brides preparing for weddings." This provides far greater diversity than simply labeling by product function or model.

[0046] Deepening the dimensionality of scenario integration: Combining the user's specific scenario (such as season, holiday, life stage, etc.) with product information will generate new and meaningful tag combinations. For example, after integrating user information, winter thermal underwear is no longer simply "winter thermal underwear" but can be refined into "professional thermal underwear suitable for ski enthusiasts in winter sports scenarios," greatly expanding the scenario-based content covered by the tag.

[0047] In summary, by integrating user information and product information into global tags, the diversity of tags can be improved, thereby making the messages to be pushed generated based on the global tags more diverse.

[0048] The specific implementation details of the embodiments of the present application are illustrated below with reference to the accompanying drawings.

[0049] The method of the embodiment of the present application can be applied to Figure 1 The server 102 in the application scenario shown. It can be understood that Figure 1 The application scenarios shown are merely illustrative of the application scenarios of this application and are not intended to limit this application.

[0050] In step S12, a tag set may be pre-generated and stored, wherein the tag set includes multiple global tags, each of which is associated with user information and product information.

[0051] User information describes user attributes, including but not limited to age, gender, occupation, user level, spending power, and frequency of use of e-commerce platform services. Product information describes product attributes, including but not limited to price, style, material, discount information, and usage. Global tags describe the association between users and products on the e-commerce platform. This association between users and products can describe shared attributes between the user and the product—that is, both the user and the product possess the attributes described by the global tag.

[0052] In some embodiments, a global tag may be generated by a large language model. The following example illustrates the process of obtaining a global tag.

[0053] like Figure 3A As shown, a user information set including multiple user information and a product information set including multiple product information can be obtained, and multiple user information in the user information set and multiple product information in the product information set are matched to obtain multiple matching information pairs, each matching information pair including at least one user information and at least one product information, and a global label is generated for each matching information pair by a large language model. In this embodiment, the user information and product information can be matched multiple times by a large language model. Each time a match is made, several user information are randomly selected from the user information set, and several product information are randomly selected from the product information set. The large language model can output a matching result, which is used to indicate whether there is matching information between the selected user information and the selected product information. Furthermore, if so, the large language model can also output the reason why the user information and product information match.

[0054] For example, if user information includes the following information: {gender: female; region: United States; product clicked by the user: women's handbags}, and product information includes the following information: {first-level category: clothing; second-level category: tops and T-shirts; third-level category: T-shirts; style: fast fashion}, the output of the large language model can be as follows:

[0055] Category label: 1;

[0056] Global labels: {Women's clothing: [0.7, consistent with user gender and interests]; Fashion matching: [0.6, related to user handbag clicks]; Casual wear: [0.7, suitable for user comfort needs]; Seasonal fashion: [0.6, clothing suitable for the current season]}. The classification label takes a value of 0 or 1, where 0 indicates that there is no matching information between the user and product information, and 1 indicates that there is a matching information between the user and product information. "Women's clothing," "Fashion matching," "Casual wear," and "Seasonal fashion" are global labels. The numerical value in brackets after the global label indicates the probability that the user and product information match within the dimension of the global label, and the text in brackets indicates the reason why the user and product information match within the dimension of the global label.

[0057] For example, if user information includes the following: {Gender: Male; Age: 45-54; Products clicked by the user: Woodworking machine tools, power tool parts and accessories}, and product information includes the following: {First-level category: Clothing; Second-level category: Tops and T-shirts; Third-level category: T-shirts; Style: Fast fashion}, the output of the large language model can be as follows:

[0058] Category label: 0;

[0059] Global tag: {}.

[0060] This means that there is no matching information between the above user information and product information.

[0061] like Figure 3B As shown, the user's behavior sequence on the e-commerce platform can be input into the large language model, so that the large language model obtains global labels based on the behavior sequence, and generates user information and product information that matches the global labels through the large language model.

[0062] The behavior sequence can include user actions (such as clicks, favorites, and purchases) on a number of products. For example, the behavior sequence might include actions on shock absorber parts, nuts and bolts, fuel filters, wipers, and parking brake parts in the Auto Parts & Accessories category. This behavior sequence can be input into a large language model, and the global label output by the large language model might be "technology enthusiast." The large language model can also generate user and product information that matches this global label. For example, user information might include "Gender: Male; Age: 30-40; Interests: Auto Repair and Maintenance, DIY Projects; Consumption Behavior: Focuses on product quality and technical specifications, and is willing to pay a higher price for high-quality auto parts; Recent Purchase History: Fuel Filters, Auto Tool Sets." Product information might include "Shock absorber parts; Nuts and bolts; Fuel filter; Wiper; Parking brake parts."

[0063] The above embodiment obtains global labels through a large language model, which can fully utilize the world knowledge learned by the large language model and improve the accuracy of label acquisition.

[0064] In some embodiments, the product information associated with the global tag information may be information obtained by binning the original product information. Binning is a data preprocessing technique used to group continuous variables or discrete data into discrete intervals or "bins." Through binning, continuous data can be converted into categorical data, making the data easier to analyze and interpret while reducing the impact of noise and outliers on the data. For example, if product information includes price, the price can be binned to obtain multiple price ranges, such as less than 100 yuan, 101-200 yuan, 201-300 yuan, and so on.

[0065] Similarly, the product information associated with global tag information can be obtained by binning the original user information. For example, if the user information includes age, the age can be binned to obtain multiple age groups, such as under 20 years old, 21-30 years old, 31-40 years old, 41-50 years old, etc.

[0066] Through binning, a transition space can be built to compress the original user information and original product information, extract common feature descriptions from the original user information and original product information, and achieve dimensionality reduction processing of the original user information and original product information.

[0067] In some embodiments, after obtaining the global tag, the association relationship between the global tag and user information can be determined through the large language model, as well as the association relationship between the global tag and product tags. However, the process of calling the large language model is time-consuming and resource-intensive. In order to improve the efficiency of obtaining the above-mentioned association relationship and reduce the resource consumption in the process of establishing the above-mentioned association relationship, the embodiments of the present application provide another method for establishing the association relationship between the global tag and user information / product information, which is explained below.

[0068] In some embodiments, sample data can be obtained, including a sample global label and sample user information and sample product information associated with the sample global label. A classification model is trained based on the sample data. The classification model is used to classify user information and product information. The category of user information is used to represent the global label associated with the user information, and the category of product information is used to represent the global label associated with the product information. Based on the classification results of the classification model, the global label is associated with the user information and the product information, respectively. The association between the sample global label and the sample user information and the sample product information is obtained through a large language model.

[0069] In this embodiment, the sample user information may include part of the full user information, and the sample product information may include part of the full product information. This embodiment uses a part of the user information and a part of the product information as sample data, obtains the association between the user information and product information in the sample data and the global label through a large language model, and trains a classification model through the sample data to determine the category of the user information and product information. Since the process of calling the classification model is shorter and occupies fewer resources than the process of calling the large language model, the above method can effectively improve efficiency and reduce resource consumption in the process of establishing the association between the global label and the user information / product information.

[0070] In some embodiments, multiple initial global labels can be obtained and pruned to obtain multiple global labels included in the label set. The initial global labels can be obtained using the method described in any of the aforementioned embodiments and will not be further described here. Pruning can remove redundant labels and improve the quality of global labels.

[0071] Specifically, multiple initial global labels can be clustered to obtain multiple cluster centers, and these multiple cluster centers are determined as multiple global labels included in the label set. Among them, the clustering algorithm can adopt K-Means, hierarchical clustering (Hierarchical Clustering) or DBSCAN and other algorithms, which are not limited in this application. Through clustering, multiple representative cluster centers can be obtained, thereby integrating multiple initial global labels into more refined and representative labels.

[0072] In step S14, a target global tag may be determined from the tag set. For example, several global tags may be randomly selected from the tag set as the target global tag. For another example, a global tag in the tag set that matches the target event may be determined as the target global tag. The target event includes, but is not limited to, any of the following:

[0073] Trending events are events whose user search volume, discussion intensity, or dissemination intensity reaches a certain threshold within a specific time period. These events are typically characterized by high attention, strong dissemination, and strong timeliness, reflecting the public's immediate interests and social hot spots. By identifying global tags matching trending events as target global tags, users can be provided with high-quality, timely information, allowing them to understand current hot and popular trends. By leveraging the appeal and influence of trending events, attractive tags and messages can be generated, effectively stimulating users' curiosity and desire to explore, increasing platform traffic and user activity.

[0074] Time-sensitive events are events that have significant value for decision-making, communication, or social impact within a specific timeframe, and their value decays over time. The core characteristic of these events is time sensitivity: the utility, influence, or social impact of the event is closely tied to the time of its occurrence. Time-sensitive events can be seasonal or festive. For example, during the summer, e-commerce platforms focus on summer-related products and user demand. Identifying global tags associated with time-sensitive events can help e-commerce platforms generate timely messages.

[0075] Events are short-term or long-term activities designed by organizers to achieve specific goals (such as cultural dissemination, brand promotion, or public welfare advocacy). These activities can take various forms, including celebrations, competitions, exhibitions, markets, and forums. During an event, e-commerce platforms can filter relevant target global tags from a tag collection based on the event theme. These tags are closely related to the event theme. By identifying these tags as target global tags, e-commerce platforms can accurately push event-related information to users, increasing their participation in the event.

[0076] In step S16, a message to be pushed corresponding to the target global tag and including the target product information is generated through the large language model.

[0077] Specifically, the prompt information including the target global tag can be input into the large language model, so that the large language model generates an initial message corresponding to the target global tag based on the prompt information, and fills the target product information into the initial message to obtain a message to be pushed corresponding to the target global tag and including the target product information.

[0078] For example, a prompt might read: "The user is a sneaker enthusiast and is currently searching for a pair of running shoes. Here are some product recommendations that match the user's preferences: {Product List}." The initial message generated by the large language model might read: "Dear sneaker enthusiasts, we recommend the following running shoes: {Product 1 Name}, lightweight and shock-absorbing, priced at {Price}; {Product 2 Name}, breathable and comfortable, priced at {Price}." By inserting the product names and prices into this initial message, a push message corresponding to the target global tag and including the target product information is generated. By combining the target global tag with the capabilities of the large language model, this method can generate more personalized, accurate, and engaging messages, improving message diversity.

[0079] In some embodiments, the prompt information also includes constraints for constraining the format of the message generated by the large language model. The format of the message includes but is not limited to the language type of the message (such as Chinese, English), the number of characters, the language style, the font, specific formatting marks (such as underscores), symbols (such as emoticons, punctuation marks) and / or sentence types (such as questions, exclamations). On this basis, the prompt information including the target global label and the constraints can be input into the large language model so that the large language model generates an initial message corresponding to the target global label and satisfying the constraints based on the prompt information.

[0080] In some embodiments, the prompt information also includes information about the number of messages generated by the large language model. The prompt information, including the target global label, constraints, and quantity information, can be input into the large language model so that the large language model generates a corresponding number of initial messages corresponding to the target global label and satisfying the constraints based on the prompt information. Different initial messages satisfy different constraints.

[0081] For example, suppose the target global tag is "Tech Enthusiasts" and the target product information is as follows: "A new wireless headset with active noise cancellation, up to 30 hours of battery life, and ambient sound mode, priced at 899 yuan." Assume the quantity is 3, indicating that three messages are expected to be pushed. The constraints for each message are as follows:

[0082] Message 1: Language: Chinese; Number of characters: No more than 120 characters; Language style: Formal and professional; Specific formatting marks: None; Sentence type: Declarative sentence;

[0083] Message 2: Language: English; Character count: No more than 150 characters; Language style: Colloquial; Symbols: Include emoticons; Sentence type: Exclamatory;

[0084] Message 3: Language: Chinese; Number of characters: No more than 180 characters; Language style: Light-hearted and humorous; Special formatting marks: Use underscores for emphasis; Sentence type: A combination of interrogative and declarative sentences.

[0085] Then, the message to be pushed generated by the large language model can be as follows:

[0086] News 1: The new wireless headphones have active noise reduction, 30 hours of battery life and ambient sound mode, designed for technology enthusiasts, priced at 899 yuan, and bring an excellent audio experience.

[0087] Message 2: Wow! This amazing wireless headphones with active noisecancellation and superlong battery life is a game-changer for tech lovers! Only $899! You won't believe your ears! [headphone symbol].

[0088] Message 3: Are you looking for wireless headphones that immerse you in the world of music? This new wireless headphone not only has amazing active noise cancellation, but also has a long battery life, all for just 899 yuan! Want to learn more? Click here to check it out! _Key_: Buy now and there's a surprise waiting for you! [Smiley face emoji]

[0089] This approach ensures that the messages generated by the large language model meet quantity requirements while ensuring that each message complies with specific format and content constraints, thereby better meeting different marketing needs and user preferences.

[0090] In some embodiments, if the product information associated with the global tag is the information obtained after the original product information is binned, the original product information corresponding to the target product information can also be determined, and a message to be pushed corresponding to the target global tag and including the original product information corresponding to the target product information is generated through a large language model. In this way, the detailed features of the product can be retained in the message to be pushed. For example, in the binning process, the price of the product may be divided into three levels of "high", "medium" and "low", but the original price information may contain specific values, such as "99 yuan", "199 yuan", etc. Including these specific price information in the message to be pushed can provide users with more accurate information and meet the needs of different users for product details.

[0091] In step S18, the message to be pushed can be sent to the target user. The target user must meet the following conditions: their user information matches the user information associated with the target global tag. For example, if the target global tag is "sneaker enthusiast," the target user can be a user with a history of browsing or purchasing sneakers. In this way, the messages received by the user are closely aligned with their personal interests and needs, enabling precise delivery of the message, thereby increasing user attention and satisfaction.

[0092] In some embodiments, if the user information associated with a global tag is obtained by binning the original user information, the user information corresponding to the target global tag can be restored to the original user information, and users whose user information matches the original user information can be identified as target users. For example, assuming the user information corresponding to the target global tag includes age, the user information obtained after binning may be "25-30 years old." This age range can be mapped to the original user information, such as 25, 26, 27, 28, 29, and 30 years old, and users aged 25, 26, ..., or 30 can be identified as target users.

[0093] Figure 4A This example shows an overall flow chart of an embodiment of the present application. Based on given global tags, user information, and product information, this embodiment uses a large language model to build a link, obtain a more personalized and fine-grained user + product + copy matching supply, and combines online performance to perform screening and optimization to improve the overall click-through rate. This embodiment mainly includes the following steps:

[0094] (1) Information preparation: Prepare a label set based on a large language model. The global label in the label set is associated with user information and product information. There are two ways to obtain the global label:

[0095] Method 1: Align user information and product information, and generate a unified global label for the aligned user information and product information through a large language model. This solution introduces the world knowledge that the large language model has learned when generating global labels. The process of generating global labels is to learn the user / product portrait based on natural language representation from the knowledge output by the large language model, and generalize the world knowledge based on the portrait description. It is a compression process from the natural language space of world knowledge to the label space. This embodiment is aimed at the large language model, and its goal is to solve the problem of efficient fitting of world knowledge. Therefore, the process of generating global labels can be approximately understood as a "low-rank decomposition" of the high-dimensional world space, that is, assuming that the users and products of the e-commerce platform have reasoning associations based on world knowledge from the perspective of the large language model, the large language model can be used to match user information and product information to obtain matching information pairs, and generate global labels based on the matching information pairs. See Figure 4B , the steps are as follows:

[0096] (i) Information Compression: Because object (user or product) information may change daily, the daily inference cost of a large language model is too high. Therefore, this embodiment builds a transition space between the object space and the natural language space to perform information compression. Common feature descriptions are extracted from the initial user information and initial product information (i.e., the initial user information and initial product information are binned), thereby reducing the dimensionality of the original high-dimensional space.

[0097] (ii) Prompt information construction: randomly select the user information of user m and the product information of product n, separate the selected user information and product information into boxes, and construct a user-product pair. mn , generate several pairs of user-item pairs according to the above method, and construct prompt information for input into the large language model. The prompt information is used to guide the large language model to judge the pair mn Is there a connection in the world knowledge space? If so, give specific reasons.

[0098] (iii) Large language model inference: All user-item pairs are input into the large language model to generate initial global labels;

[0099] (iv) Label pruning: pruning the initial global labels to obtain the global labels in the label set;

[0100] (v) Object information restoration: The user information and product information after the binning process are restored to the initial user information and initial product information, and associated with the global tag.

[0101] (vi) Classification model training: Based on the sampled user information, product information, and global labels, a classification model is trained to obtain a classification model for classifying the input information (i.e., user information or product information). The category of the input information is used to represent the global label associated with the input information. The classification model is used to determine the associated global label for all user information and product information.

[0102] (vii) Large language model evaluation: Randomly extract user-item pairs and obtain the global labels associated with the user information and item information in the user-item pairs. Use the large language model to verify the correlation between the user information, item information and the global labels.

[0103] Method 2: A large language model generates global labels based on product information and associates the global labels with user information. Figure 4C , the steps are as follows:

[0104] (i) Prompt Information Generation: Based on multi-channel information sources, such as predefined general world knowledge that can be used for label production and behavioral e-commerce knowledge extracted from e-commerce platforms (i.e., user behavior sequences on e-commerce platforms), initial prompt information is constructed. After prompt engineering optimization and combined with the ICL\COL module, the initial prompt information is converted into prompt information suitable for the large model;

[0105] (ii) Large language model inference: Generate initial global labels through large language model inference;

[0106] (iii) Label post-processing: clustering the initial global labels through selected strategies such as kmeans clustering or semi-supervised clustering methods to improve label quality and generate a certain number of global labels required;

[0107] (iv) User / product information construction: Obtain user information and product information from the e-commerce platform.

[0108] (v) User / product information compression: Since object data may change daily, the daily inference cost of a large model is too high. Therefore, this embodiment builds a transition space between the e-commerce object space and the natural language space to perform information compression. Common feature descriptions are extracted from the initial user information and initial product information (i.e., the initial user information and initial product information are binned), thereby achieving dimensionality reduction processing on the original high-dimensional space.

[0109] (vi) Introduction of tag library information of object profiling module: The global tags generated in step (iii) are introduced into the prompt information. All global tags can be introduced, or relevant global tags can be introduced through RAG.

[0110] (vii) Prompt Engineering: The information obtained in steps (iv) to (vi) is organized into prompt information, and input is optimized in conjunction with prompt engineering.

[0111] (viii) Large language model reasoning: Generate user information and product information that matches global labels through a large language model.

[0112] (ix) Label Post-Processing: Considering that large language models are generative models, they may have inaccurate understanding or output when handling complex tasks. This may result in output labels that are similar in meaning to global labels in the label set but are not in the label set. Therefore, label post-processing is required to match and map the labels generated by the large language model with the global labels in the label set. To ensure that the final generated labels are accurately matched to the label set, this matching mapping can be achieved using text representation similarity.

[0113] (x) Object space restoration: When generating the object information of each object, it is necessary to restore the object information (user information or product information) to the original object information through the characteristics of the transition space. This can support the change of object characteristics due to daily changes, and map the latest object information from the transition space.

[0114] (xi) Functional evaluation: In the actual evaluation iteration, the generated labels are evaluated for coverage and timeliness.

[0115] (xii) Efficiency evaluation: In real scenarios, the efficiency of labeling is evaluated through the actual indicators of the scenario.

[0116] (2) Large language model generates multi-language push messages: For each global tag, 100 products are sampled, and product information such as product titles is obtained. The product information is combined with the global tag to generate prompt information. Combined with the multi-language message production requirements (i.e., the constraints in the aforementioned embodiment), the large language model is used to generate push messages in multiple languages.

[0117] (3) Risk control filtering: Conduct risk control review on the messages to be pushed, and filter out messages with risk control issues;

[0118] (4) The large language model performs preliminary screening on the messages to be pushed: the messages to be pushed generated in step (3) are spliced ​​with the original input information (product information such as product title), and the large language model is used to determine again whether the generated messages to be pushed are applicable to all the products contained therein, and whether they are suitable for being pushed to users as a topic (for example, medical topics are not suitable for push); for multilingual messages to be pushed, multiple different versions of the large language model are used to determine again whether the content of the messages to be pushed is fluent and correct, and the correct messages to be pushed are screened out by voting on the judgment results of multiple versions of the large language model.

[0119] (5) Manual review: Before the pass rate of the messages to be pushed produced by the large language model reaches the standard, a sample of products under each global label will be manually reviewed to confirm whether the label is suitable for transmission.

[0120] (6) Review of multi-language messages to be pushed: Before the pass rate of the messages to be pushed produced by the large language model reaches the standard, a sample of the multi-language messages to be pushed is manually reviewed.

[0121] (7) Message assembly: Assemble the approved global tags and corresponding product information to produce candidate tasks.

[0122] (8) Random distribution: The assembled candidate tasks will be randomly distributed.

[0123] (9) Preferred online: After a period of random distribution, the top K click-through rate messages to be pushed are screened out, and the screened messages to be pushed are distributed in experimental buckets, that is, distributed to target users in a targeted manner.

[0124] (10) Feedback optimization: Based on the targeted distribution results of step (9) and manual review feedback, guide the optimization of the output results of the aforementioned large language models.

[0125] This application has the following technical effects:

[0126] (1) A link for message push based on global tags generated by a large language model was established, which improved the diversity of messages;

[0127] (2) Use multiple large language model assistants to process specific tasks (production, review), improve the accuracy of manual review, and reduce manual review resources;

[0128] (3) Continuous iteration can be achieved by selecting the optimal tasks and optimizing production tasks based on real traffic data.

[0129] An embodiment of the present application further provides a computer device, which comprises at least a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any of the aforementioned embodiments is implemented.

[0130] Figure 5 2 shows a more specific hardware structure diagram of a computer device provided in an embodiment of the present application. The device may include: a processor 202, a memory 204, an input / output interface 206, a communication interface 208, and a bus 210. The processor 202, the memory 204, the input / output interface 206, and the communication interface 208 are connected to each other within the device via the bus 210.

[0131] The processor 202 can be implemented using a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute relevant programs to implement the technical solutions provided in the embodiments of the present application. The processor 202 can also include a graphics card, which can be an Nvidia Titan X graphics card or an 1080Ti graphics card.

[0132] The memory 204 can be implemented in the form of a read-only memory (ROM), a random access memory (RAM), a static storage device, a dynamic storage device, etc. The memory 204 can store an operating system and other application programs. When the technical solutions provided in the embodiments of the present application are implemented through software or firmware, the relevant program codes are stored in the memory 204 and are called and executed by the processor 202.

[0133] The input / output interface 206 is used to connect to input / output modules to implement information input and output. The input / output modules can be configured as components within the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0134] The communication interface 208 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, Wi-Fi, Bluetooth, etc.).

[0135] The bus 210 comprises a pathway for transmitting information between the various components of the device, such as the processor 202 , the memory 204 , the input / output interface 206 , and the communication interface 208 .

[0136] It should be noted that although the above device only shows the processor 202, memory 204, input / output interface 206, communication interface 208, and bus 210, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may only include the components necessary to implement the embodiments of the present application, and does not necessarily include all the components shown in the figure.

[0137] An embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any embodiment of the present application when executed by a processor.

[0138] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the aforementioned embodiments.

[0139] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information using 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 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, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic 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 computer device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0140] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely illustrative, wherein the modules described as separate components may or may not be physically separated, and the functions of each module can be implemented in the same one or more software and / or hardware when implementing the embodiment of this application. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the embodiment. Those of ordinary skill in the art can understand and implement it without paying any creative work.

[0141] The above is only a specific implementation of the embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the embodiment of the present application. These improvements and modifications should also be regarded as the scope of protection of the embodiment of the present application.

Claims

1. A message push method, the method comprising: Obtain a pre-generated tag set, wherein the tag set includes multiple global tags, wherein the global tags are respectively associated with user information and product information, and the global tags are used to describe association information between users of the e-commerce platform and products of the e-commerce platform; Determine a target global tag from the tag set, and obtain target product information associated with the target global tag; Generate a message to be pushed corresponding to the target global tag and including the target product information through a large language model; The message to be pushed is pushed to a target user, wherein the user information of the target user matches the user information associated with the target global tag.

2. The method according to claim 1, further comprising: Acquire a user information set including multiple user information and a product information set including multiple product information; Matching the plurality of user information in the user information set with the plurality of product information in the product information set to obtain a plurality of matching information pairs, each matching information pair including at least one user information and at least one product information; A global label is generated for each matching information pair through a large language model.

3. The method according to claim 1, further comprising: Inputting a user's behavior sequence on the e-commerce platform into a large language model, so that the large language model obtains a global label based on the behavior sequence; User information and product information matching the global tag are generated through a large language model.

4. The method according to claim 1, wherein the product information associated with the global tag information is information obtained by binning the original product information; The message to be pushed corresponding to the target global tag and including the target product information is generated by the large language model, including: Determining original product information corresponding to the target product information; Generate a message to be pushed corresponding to the target global tag and including the original product information corresponding to the target product information through a large language model; and / or The user information associated with the global tag information is information obtained by binning the original user information. The method further includes: Restoring the user information corresponding to the target global tag to the original user information; A user whose user information matches the original user information is determined as a target user.

5. The method according to claim 1, further comprising: Acquire sample data, where the sample data includes a sample global tag and sample user information and sample product information associated with the sample global tag; The association between the sample global label and the sample user information and sample product information is obtained through a large language model; Training a classification model based on the sample data; the classification model is used to classify user information and product information, the category of user information is used to represent the global label associated with the user information, and the category of product information is used to represent the global label associated with the product information; The global tag is associated with user information and product information respectively based on the classification result of the classification model.

6. The method according to claim 1, wherein generating a message to be pushed corresponding to the target global tag and including the target product information through a large language model comprises: inputting prompt information including the target global tag into a large language model, so that the large language model generates an initial message corresponding to the target global tag based on the prompt information; The target product information is filled into the initial message to obtain a message to be pushed that corresponds to the target global tag and includes the target product information.

7. The method according to claim 6, wherein the prompt information further includes a constraint condition for constraining the format of the message generated by the large language model; Inputting the prompt information including the target global tag into the large language model so that the large language model generates an initial message corresponding to the target global tag based on the prompt information includes: The prompt information including the target global label and the constraint condition is input into the large language model, so that the large language model generates an initial message corresponding to the target global label and satisfying the constraint condition based on the prompt information.

8. The method according to claim 7, wherein the prompt information further includes information on the number of messages generated by the large language model; Inputting the prompt information including the target global tag into the large language model so that the large language model generates an initial message corresponding to the target global tag based on the prompt information includes: Inputting prompt information including the target global label, the constraint conditions, and the quantity information into a large language model, so that the large language model generates a corresponding number of initial messages corresponding to the target global label and satisfying the constraint conditions based on the prompt information; wherein different initial messages satisfy different constraint conditions.

9. The method according to claim 1, wherein determining a target global tag from the tag set comprises: Get the target event; The target event includes any of the following: hot search events, time-sensitive events, and event events; A global tag in the tag set that matches the target event is determined as a target global tag.

10. The method according to claim 1, further comprising: Get multiple initial global labels; Pruning is performed on the multiple initial global labels to obtain multiple global labels included in the label set.

11. The method according to claim 10, wherein pruning the multiple initial global labels to obtain the multiple global labels included in the label set comprises: Clustering the multiple initial global labels to obtain multiple cluster centers; The multiple cluster centers are determined as a multiple global tags included in the tag set.

12. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 11 when executing the computer program.

14. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images