Message sending method and device based on natural language processing, equipment and medium

Abstract

The application relates to a message sending method and device based on natural language processing, equipment and a storage medium. The method comprises the following steps: analyzing an unstructured text instruction input by an operator to obtain a strategy parameter, constructing a user multi-channel identity graph and aggregating a push frequency and a reach load, evaluating a channel expected income based on a structured basic strategy parameter, a multi-channel identity relationship graph, the push frequency and the reach load to select a reach path and determine an intercepted user, dividing a sending window into dynamic execution slots, mapping the user to a target slot according to a user active preference to obtain a target correspondence relationship between the user and the target slot, generating an individualized marketing script through a natural language processing model, assembling the individualized marketing script, the reach path and the target correspondence relationship into a sending task after excluding the intercepted user, and sending a message corresponding to the sending task to a corresponding channel. The application realizes global scheduling control of cross-application marketing message sending.

Description

Technical Field

[0001] This application relates to the field of message push technology, and in particular to a message sending method, apparatus, device and storage medium based on natural language processing. Background Technology

[0002] With the deepening development of the mobile internet, enterprises typically operate multiple applications to build an ecosystem of services. In actual operation, each application is often managed by an independent business team, each maintaining its own user outreach channels and executing its own marketing messaging strategies.

[0003] In existing technologies, the sending of marketing messages mainly relies on fixed logic, such as rules for automatically resending SMS messages if a message is missed, preset static template text, and sending limits within a single application. Because user data across applications is isolated, when sending a marketing message to a user in one application, the system cannot obtain the number and frequency of messages the user has received in other applications, or the message backlog in the current sending queue. This results in each application making its sending decisions based on its own independent and incomplete data view. Without a mechanism to perceive the global reach status, when different applications initiate sending requests to the same user at similar times, it is impossible to assess the incremental impact of each sending request on the user's total reach load before making a single sending decision, nor can it perform unified timing planning and conflict coordination for multiple sending requests over time. Furthermore, existing technologies use fixed static templates at the text generation level, and the text content is not adjusted according to the individual characteristics of the target user or the format constraints of the selected sending channel.

[0004] Therefore, in scenarios where multiple applications independently initiate sending requests, how to filter, schedule, and adapt sending requests based on global reach status to achieve unified scheduling and control across applications has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above, this application provides a message sending method, apparatus, device and storage medium based on natural language processing, the purpose of which is to solve the above-mentioned technical problems.

[0006] Firstly, this application provides a message sending method based on natural language processing, the method comprising: The unstructured text instructions input by operators are parsed using a natural language processing model to obtain structured basic strategy parameters; Construct a user identity relationship graph with multiple channels, and aggregate the push frequency and reach load of users within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load, the expected benefits of multiple channel modes are evaluated to select the reach path and identify the blocked users. The sending window is divided into multiple dynamic execution slots. Users are mapped to target slots based on their historical active time preferences, thus obtaining the target correspondence between users and target slots. Based on user profile tags and the channel characteristics corresponding to the reach path, personalized marketing copy is generated through a natural language processing model. The personalized marketing copy, the reach path, and the target correspondence are combined into a sending task after excluding the intercepted users, and the message corresponding to the sending task is sent to the corresponding channel.

[0007] Secondly, this application provides a message sending device based on natural language processing, the message sending device based on natural language processing includes: Parsing module: Used to parse unstructured text instructions input by operators based on natural language processing models to obtain structured basic strategy parameters; Construction module: Used to construct the user's identity relationship graph with multiple channels, and aggregate the user's push frequency and reach load within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Evaluation module: used to evaluate the expected benefits of multiple channel modes based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load in order to select the reach path and identify the blocked users; Mapping module: used to divide the sending window into multiple dynamic execution slots, and map users to target slots according to users' historical active time preferences, so as to obtain the target correspondence between users and target slots; Generation module: used to generate personalized marketing copy based on user profile tags and the channel characteristics corresponding to the reach path through a natural language processing model; Sending module: Used to assemble the personalized marketing copy, the reach path and the target correspondence, excluding the blocked users, into a sending task, and send the message corresponding to the sending task to the corresponding channel.

[0008] Thirdly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the message sending method based on natural language processing as described in any embodiment of the first aspect.

[0009] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the message sending method based on natural language processing as described in any embodiment of the first aspect.

[0010] The technical solutions provided in this application have the following advantages compared with the prior art: By constructing a user identity relationship graph across multiple channels and aggregating user push frequency and reach load within the application matrix, the system can obtain the global reach status of users across all applications and channels, enabling it to have global awareness before making sending decisions. By identifying and excluding intercepted users during task assembly, the number of sending requests sent to the same user per unit time is directly reduced at the scheduling level. Simultaneously, by dividing the sending window into multiple dynamic execution slots and mapping users to target slots based on their historical active time preferences, sending requests are distributed across time periods that match the user's actual receiveability, preventing multiple sending requests from crowding out the same user's receive time window. Furthermore, personalized marketing copy is generated based on the channel characteristics of the reach path and user profiles, ensuring that the format and style of the delivered content are adapted to the target channels and user characteristics. This achieves global scheduling control over independent sending requests from multiple applications. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a preferred embodiment of the message sending method based on natural language processing in this application; Figure 2 This is a schematic diagram of a preferred embodiment of the message sending device based on natural language processing in this application; Figure 3 This is a schematic diagram of a preferred embodiment of the electronic device of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0015] It should be noted that the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed in this application.

[0016] Reference Figure 1 The diagram shown is a flowchart illustrating an embodiment of the message sending method based on natural language processing according to this application. The method is executed by an electronic device, which can be implemented by a software system and / or a hardware system. The message sending method based on natural language processing includes: Step S10: Parse the unstructured text instructions input by the operators based on the natural language processing model to obtain the structured basic strategy parameters; Step S50: Construct a user identity relationship graph with multiple channels, and aggregate the push frequency and reach load of the user within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Step S60: Based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load, evaluate the expected benefits of multiple channel modes to select the reach path and identify the blocked users; Step S40: Divide the sending window into multiple dynamic execution slots, map users to target slots according to their historical active time preferences, and obtain the target correspondence between users and target slots; Step S50: Based on user profile tags and the channel characteristics corresponding to the reach path, generate personalized marketing copy using a natural language processing model; Step S60: Assemble the personalized marketing copy, the reach path, and the target correspondence, excluding the intercepted users, into a sending task, and send the message corresponding to the sending task to the corresponding channel.

[0017] In this embodiment, the system receives unstructured text instructions input by operations personnel through an interactive interface, and uses a natural language processing model to parse the instructions, generating structured basic strategy parameters. For example, if an operations person inputs "Issue 20% off coupons to users who have browsed the site in the last three days but haven't placed an order, with the issuance time no later than 10 PM," the system receives this unstructured text instruction and inputs it into a pre-optimized natural language processing model. This model performs semantic parsing on the text, extracting key elements such as marketing objectives, cost preferences, and channel constraints. Marketing objectives refer to the description of the target user group, cost preferences refer to the acceptable range of costs for a single marketing campaign, and channel constraints refer to the restrictions on specified distribution channels. The extracted elements are then encapsulated into structured basic strategy parameters according to a preset field format. For the example instruction above, the generated structured basic strategy parameters include the marketing objective as "users who have browsed the site in the last three days but haven't placed an order," the content as "20% off coupon," the expiration time as "10 PM," and the channel preference as "none."

[0018] After obtaining the structured basic strategy parameters, a user identity relationship graph with multiple channels is constructed, and global push frequency and reach load are aggregated. In a multi-application matrix scenario, the same user may have different account identifiers in different applications and may be associated with multiple mobile phone numbers for SMS services. Therefore, the system first obtains each user's identity identifier in each application and contact information in each SMS service to establish the association between user identity and multiple channels. Specifically, the system integrates user information from all applications under the enterprise through a user data platform, using the user's natural person dimension as the unique primary key, mapping the same user's identifiers in different applications and the contact information corresponding to different SMS services to the same user entity, forming the user's identity association with multiple channels. Based on this association, the system calculates in real time the user's historical push frequency and reach load in each business line within the application matrix. Push frequency refers to the number of marketing messages received by the user per unit time, and reach load refers to the number of pending messages to be sent that the user has not yet processed.

[0019] The system constructs a multi-channel identity relationship graph by linking users to their identities across multiple channels. It then aggregates the push frequency and reach load of users across each business line into global push frequency and reach load. The resulting multi-channel identity relationship graph allows for quick lookup of a user's identity mapping across all channels, and the push frequency and reach load reflect the marketing pressure a user has experienced within the current time window. For example, if a user has received three push notifications in application A within the past hour and has one pending message in application B, the aggregated push frequency would be the three previously sent messages, and the reach load would be the one pending message.

[0020] The system evaluates the expected returns of various channel models based on structured basic strategy parameters, multi-channel identity relationship graphs, push frequency, and reach load to select the reach path for each user and simultaneously identify blocked users. Specifically, the structured basic strategy parameters, the user's multi-channel identity relationship graph, push frequency, and reach load are input into a pre-trained multi-objective machine learning model. This model predicts the expected conversion probability and expected delivery cost for each user. The expected conversion probability refers to the likelihood that the user will perform the target behavior after receiving a marketing message, and the expected delivery cost refers to the resource cost required to complete the delivery.

[0021] The system iterates through various predefined channel modes, including application push notifications, SMS services, in-app messages, and combinations thereof. For each channel mode, the system calculates the expected return based on the user's projected conversion probability and projected delivery cost. The expected return is calculated by multiplying the projected conversion probability by a preset conversion value and subtracting the projected delivery cost. The system selects the channel mode that maximizes the expected return as the user's reach path, thus determining a personalized optimal reach path for each user. Simultaneously, based on multi-channel identity relationship graphs, push frequency, and reach load, the system calculates the user's churn risk value in real time. The churn risk value refers to the probability that the user will disable notification permissions or uninstall the application due to excessive intrusion. When the calculated churn risk value exceeds the system's preset safety threshold, the user is marked as a blocked user. This process selects the most profitable reach path for each user and proactively identifies users who may churn due to excessive marketing.

[0022] Because the timing of marketing message delivery significantly impacts conversion rates, messages sent during periods of low user activity are often ignored. Therefore, the daily delivery window is first divided into multiple dynamic execution slots based on hourly granularity and channel mode. Each dynamic execution slot corresponds to a specific hourly time period and a specific channel mode. Then, each user's historical active time preference is retrieved from the user data platform. This preference is derived from statistical analysis of the time distribution of users opening applications or replying to messages over a past period, representing the time interval during which users are most likely to read messages. Finally, based on each user's historical active time preference, the user is mapped to an optimal slot among the multiple dynamic execution slots. The mapping rule is to select the slot whose time period best matches the user's active time preference, while ensuring that the channel mode of that slot is consistent with the previously selected reach path for that user. For example, if a user's active time preference is 9 PM, and the previously selected reach path for them was application push notifications, then the user is mapped to a dynamic execution slot for the 9 PM to 10 PM time period with application push notifications as the channel mode. After mapping is completed, the system establishes a correspondence between user identifiers and optimal slot identifiers. This correspondence ensures that each user's marketing messages will be sent in the most appropriate channel mode during the time period when they are most likely to read them.

[0023] After obtaining the correspondence between users and optimal slots, personalized marketing copy is generated using a natural language processing model based on user profile tags and the channel characteristics corresponding to the reach paths obtained in the previous steps. This generates personalized content for each user that matches their individual characteristics and channel characteristics. Specifically, user profile tags for each user are obtained from the user profile database. These tags include structured information such as the user's demographic characteristics, consumption preferences, and historical behavior. Simultaneously, the channel characteristics corresponding to the reach path selected for that user are obtained. Channel characteristics refer to the format restrictions and style requirements of different sending channels; for example, app push notifications support rich text, SMS services have character limits and typically use concise expressions, while in-app messages support longer content.

[0024] Next, each user's user profile tags and channel characteristics are input into the copy generation interface of the natural language processing model. The model generates copy containing relevant keywords based on the consumption preferences in the user profile tags, while adjusting the copy length and tone according to channel characteristics. After the model outputs candidate marketing copy, the best copy is selected as the personalized marketing copy for that user according to preset rules. For example, for a female user who prefers beauty products, if the delivery path is an app push notification, the model might generate personalized copy containing phrases like "Your favorite foundation is on sale for a limited time." If the delivery path is a text message service, the model generates a more concise short copy. This ensures that the marketing copy received by the user is no longer templated content, but rather customized content that integrates personal interests and channel characteristics.

[0025] Finally, the personalized marketing copy, the reach path for each user, the target correspondence between users and optimal slots, and the blocked users are comprehensively processed and assembled into a sending task, which is then distributed to the corresponding channels. Specifically, all the above data is collected to clarify the personalized marketing copy, reach path, and optimal slot for each user. Then, all entries corresponding to blocked users are deleted from the target correspondence, as blocked users have a high risk of churn and should not receive this marketing message again. After the deletion operation, the effective correspondence is obtained, which only includes unblocked users and their respective optimal slots. Next, the personalized marketing copy, reach path, and effective correspondence are assembled according to the data structure of the sending task. One sending task is assembled into an independent data packet, which contains a list of users to be sent to, the personalized marketing copy for each user, the reach path for each user, and the sending time in the optimal slot for each user.

[0026] After assembly, the sending task is dispatched to the message sending unit. Upon receiving the task, the message sending unit, based on the sending time specified by the optimal slot in each record, sends the corresponding personalized marketing copy to the designated channel according to the corresponding reach path. For example, for a user in a sending task, whose personalized marketing copy is a beauty discount ad, the reach path is app push notification, and the optimal slot corresponds to 21:00, the sending module will send the copy to the user's in-app notification bar at 21:00 via the app push service. After dispatching, the reach data will be fed back to the data center for updating user profiles and training decision-making models.

[0027] In one embodiment, the process of parsing unstructured text instructions input by operators using a natural language processing model to obtain structured basic strategy parameters includes: Input unstructured text instructions from operations personnel into a natural language processing model; The unstructured text instructions are segmented and intent is identified based on the natural language processing model to obtain segmentation results and intent identification results. Based on the word segmentation results and the intent recognition results, marketing objectives, cost preferences, and channel constraints are extracted from the unstructured text instructions; The marketing objectives, cost preferences, and channel constraints are combined into structured basic strategy parameters.

[0028] Operators input unstructured text commands into the interactive interface. These commands are descriptions in everyday language without any formatting constraints. The system takes this raw text as input and directly passes it to a pre-trained natural language processing model, without requiring any manual preprocessing or field splitting.

[0029] The natural language processing model performs word segmentation and intent recognition on the received unstructured text instructions. Word segmentation involves dividing a continuous text sequence into independent word units. For example, the instruction above is segmented into word units such as "give," "last three days," "browsed," "no order placed," "of," "user," "issue," "20% off coupon," "issuance time," and "no later than." Intent recognition, based on word segmentation, determines the overall marketing action type that the operator intends to perform. After fine-tuning with a large amount of marketing instruction data, the model can identify common intent types such as "issue coupons," "send notifications," and "create activities." For the above instruction, the model identifies the intent as "create a coupon issuance task." After these two operations, the model outputs the word segmentation result and the intent recognition result.

[0030] Based on word segmentation and intent recognition results, the model extracts three key elements from unstructured text instructions: marketing objectives, cost preferences, and channel constraints. Marketing objectives refer to the description of the target user group. The model extracts this element by identifying the conditional phrase "users who have browsed but not placed an order in the last three days." Cost preferences refer to the acceptable range of cost for a single marketing campaign. If the instruction does not explicitly specify this (e.g., the absence of keywords like "cost" or "expense" in this example), the model uses preset default values, such as no cost limit or using the system's default threshold. Channel constraints refer to the restrictions on sending channels specified by the operator, such as "only through app push notifications" or "do not send SMS messages." In the above instruction, since no channel restrictions are mentioned, the model sets the channel constraint to "none" or "any channel." If the instruction includes "do not send SMS messages," the model will extract the constraint "prohibit SMS service."

[0031] The system extracts marketing objectives, cost preferences, and channel constraints, and concatenates them into structured basic strategy parameters according to a preset data format. The concatenation method can be to combine the three fields as key-value pairs. For example, the marketing objective could be "users who have browsed but not placed an order in the last three days," the cost preference "default," and the channel constraint "none." Additionally, the explicitly given "20% off coupon" is included as a content field, and "22:00" is included as an expiration time field. The final structured basic strategy parameters form a structured data object. This transforms natural language descriptions into machine-understandable and computable structured parameters, realizing the conversion from operator intent to system instructions.

[0032] In one embodiment, constructing a user-multi-channel identity relationship graph and aggregating the user's push frequency and reach load within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load includes: Obtain the user's identity identifier in various applications and SMS services, and establish the identity association relationship between the user and multiple channels; Based on the user's identity association with multiple channels, the push frequency and reach load of the user in each business line within the application matrix are statistically analyzed to obtain the user's push frequency and reach load in each business line. The user's identity association with multiple channels is constructed into a multi-channel identity relationship graph; The push frequency and reach load of the user in each business line are aggregated into the push frequency and reach load of the user in the application matrix, thus obtaining the user's multi-channel identity relationship graph, push frequency and reach load.

[0033] The system retrieves each user's login account, device identifier, and authorized mobile phone number for each SMS service from the user data platform. These identifiers are linked using the user's natural person as a unique primary key. For example, the user ID in application A, the open identifier in application B, and the mobile phone number bound to the SMS service gateway are mapped to the same user entity, thereby establishing an identity association between the user and multiple channels. This relationship clarifies the user's representation across all applications and all SMS services.

[0034] Based on the established user and multi-channel identity associations, the system further calculates the push frequency and reach load for each user across all business lines within the application matrix. Business lines include the push channels of each application and each SMS service channel. Push frequency refers to the number of marketing messages the user has received from a particular business line within a preset time window (e.g., the past hour), while reach load refers to the number of unprocessed or unsent messages the user has at the current moment within that business line. By traversing the user's historical sending records and current pending messages across all business lines, the system obtains the user's push frequency and reach load for each business line. For example, for user Zhang San, the system calculates a push frequency of three times per hour and a reach load of one unsent message on application A's push channel; and a push frequency of zero times per hour and a reach load of zero unsent messages on the SMS service channel.

[0035] The user's identity association with multiple channels is then explicitly constructed into a multi-channel identity relationship graph. This graph uses users as nodes, with each application and SMS service as adjacent edges, and each edge is labeled with the user's corresponding identity identifier. The graph can be constructed using a graph database, for example, using the user's unique identifier as the vertex and channel type and in-channel identifier as attributes, forming a structured relationship graph that can be quickly queried. This graph is used to query in real time whether a user has a cross-application identity mapping.

[0036] The system aggregates the push frequency and reach load collected from each business line. The aggregation operation adds up the push frequencies of the same user across different business lines to obtain the user's total push frequency within the application matrix. It also adds up the reach loads across different business lines to obtain the user's total reach load within the application matrix. For example, if user Zhang San's push frequency is three times per hour and reach load is one message in application A's push channel, twice per hour and zero messages in application B's push channel, and zero times per hour and zero messages in the SMS service channel, then the aggregated global push frequency is five times per hour and the global reach load is one message.

[0037] In one embodiment, the step of evaluating the expected returns of multiple channel modes to select an outreach path and identify blocked users based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load includes: The structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load are input into a multi-objective machine learning model to obtain the expected conversion probability and expected delivery cost for each user. Based on each user's expected conversion probability and expected delivery cost, calculate each user's expected revenue under multiple channel models, and select the reach path with the highest expected revenue for each user to obtain each user's reach path; churn risk value for each user is calculated based on the multi-channel identity relationship graph, the push frequency, and the reach load. Users whose churn risk value exceeds a preset security threshold are marked as blocked users.

[0038] The model inputs structured basic strategy parameters, the user's multi-channel identity relationship graph, push frequency, and reach load into a pre-trained multi-objective machine learning model. This model employs a multi-task learning architecture, simultaneously outputting predictions for both the expected conversion probability and the expected delivery cost. The expected conversion probability refers to the likelihood that a user will perform a target action (such as placing an order or claiming a coupon) if a marketing message is sent to them; it is typically represented by a value between 0 and 1. The expected delivery cost refers to the resource costs incurred in completing the delivery, such as SMS channel fees and push service call fees. During the training phase, the model uses historical marketing data, including labels such as user characteristics, sending channels, sending time, conversion rate, and actual cost, enabling the model to extract effective features from the input multi-channel identity relationship graph, push frequency, and reach load. For example, for user Zhang San, the model outputs an expected conversion probability of 0.75 and an expected delivery cost of 0.05 yuan.

[0039] Based on each user's estimated conversion probability and estimated delivery cost, the expected revenue for that user across multiple channel modes is calculated. Predefined channel modes include four types: app-only push notifications, app-first followed by SMS service, SMS service only, and in-app messaging. For each channel mode, a pre-configured conversion value parameter (e.g., average revenue per conversion) is used to calculate the expected revenue. The formula is: Expected Revenue = Estimated Conversion Probability × Conversion Value - Estimated Delivery Cost.

[0040] The expected delivery cost can vary depending on the channel model. For example, the delivery cost of SMS service only is higher than that of application push only. However, the expected conversion probability may also vary depending on the channel characteristics. The multi-objective machine learning model has already output the corresponding expected conversion probability and expected delivery cost for each channel model. It iterates through all channel models and selects the one that maximizes the expected return as the user's reach path. Taking user Zhang San as an example, if the expected return for application push only is 0.8 yuan, the expected return for application first followed by SMS service is 0.6 yuan, the expected return for SMS service only is 0.3 yuan, and the expected return for in-app message is 0.2 yuan, then the system selects application push only as Zhang San's reach path. A personalized optimal reach path can be generated for each user.

[0041] Then, based on the multi-channel identity relationship graph, push frequency, and reach load, the churn risk value for each user is calculated. The churn risk value refers to the probability that a user will uninstall the application or disable notification permissions due to excessive disturbance. The calculation of the churn risk value can employ a logistic regression model or a gradient boosting tree-based model. The model's input features include the user's total push frequency within the application matrix (i.e., the total number of marketing messages received per unit time), the cumulative reach load of each business line, the recent trend of the user's click-through rate (e.g., whether the click-through rate has continuously decreased over the past seven days), and the number of days since the user last actively opened the application. The model outputs a risk score between 0 and 1. For example, for user Zhang San, based on his multi-channel identity relationship graph, he is registered in three applications and has received five push notifications in the past hour, with a reach load of two pending messages. His click-through rate has decreased from 20% to 5% in the past three days. The model calculates his churn risk value to be 0.82.

[0042] The system marks users whose churn risk value exceeds a preset safety threshold as blocked users. This threshold is dynamically set by operations personnel based on overall user experience goals and marketing budgets, for example, 0.7. Each user's churn risk value is compared to the threshold. If the churn risk value is greater than or equal to the threshold, the user is marked as blocked; otherwise, no mark is made. Using the example above, user Zhang San's churn risk value is 0.82, exceeding the threshold of 0.7. Therefore, Zhang San is marked as a blocked user and will be excluded from the sending queue, thus avoiding sending marketing messages to users already at high churn risk and reducing user resentment. Users with a churn risk value of 0.2 will not be marked and can continue to participate in subsequent sending processes.

[0043] In one embodiment, dividing the sending window into multiple dynamic execution slots and mapping users to target slots based on their historical active time preferences to obtain the target correspondence between users and target slots includes: The sending window is divided into multiple dynamic execution slots based on hourly granularity and the number of channel modes; Obtain each user's historical active time preference, and map each user to a target slot among the multiple dynamic execution slots based on the user's historical active time preference, thereby obtaining the target correspondence between the user and the target slot.

[0044] The system divides the daily sending window into hourly granularity and the number of channel modes. A sending window refers to the entire day's time range during which marketing messages can be sent, typically 24 hours. The system further divides the 24 hours into 24 hourly segments, each corresponding to a specific hourly interval. Simultaneously, the system retrieves the number of preset channel mode types, such as four modes: application-only push, application-first followed by SMS service, SMS service only, and in-app messaging. The 24 hourly segments and the four channel modes are combined using a Cartesian product to form 96 dynamic execution slots. Each dynamic execution slot has a unique identifier and is associated with both the hourly segment and the channel mode. These 96 dynamic execution slots cover all possible sending opportunities and channel combinations.

[0045] The system retrieves each user's historical active time preference from the user data platform. This preference is derived from statistical analysis of the time distribution of a user's app openings, push notification clicks, and SMS replies over a given period (e.g., the past 30 days). Specifically, the system counts the frequency of user activity on an hourly basis, generating a 24-dimensional activity vector, where each dimension represents the number of active events within that hour. After normalization, the one or more consecutive hourly segments with the highest activity level represent the user's historical active time preference. For example, if a user opened the app 80% of the time between 8 PM and 10 PM each night over the past 30 days, their historical active time preference would be 8 PM to 10 PM.

[0046] Based on each user's historical active time preferences, the system maps each user to a target slot from 96 dynamic execution slots. The mapping rule can be to filter out a set of slots from all dynamic execution slots whose channel mode matches the user's selected reach path, and then select the slot with the highest match between the time period and the user's historical active time preferences from this set as the target slot. The match degree can be measured by the absolute difference between the midpoint of the time period and the center time of the user's preferred time period; the smaller the difference, the higher the match degree. If multiple slots have the same match degree, they can be selected randomly or in order of slot number. For example, if a user's historical active time preference is 21:00, and their reach path is "application-only push," then the system selects the slot whose time period includes 21:00 or is closest to 21:00 from all dynamic execution slots with the channel mode "application-only push," as the user's target slot.

[0047] The system records the correspondence between each user and their assigned target slot, forming a target-to-target correspondence. This correspondence can be stored as a list of key-value pairs, where the key is the user identifier and the value is the unique identifier of the target slot. For example, user Zhang San corresponds to slot number 57 (9 PM to 10 PM, application-only push notifications), and user Li Si corresponds to slot number 12 (4 PM to 5 PM, SMS service). The target correspondence is used to determine the specific sending time and channel mode for each user, achieving an improvement from a coarse-grained, unified sending approach to a refined, individual time-slot scheduling approach, ensuring that each user's marketing messages are delivered at the most suitable time possible.

[0048] In one embodiment, generating personalized marketing copy based on user profile tags and the channel characteristics corresponding to the reach path using a natural language processing model includes: Obtain user profile tags for each user and channel characteristics corresponding to the selected reach path for each user; Each user's user profile tags and channel characteristics are input into a natural language processing model. The natural language processing model then generates copywriting content that matches the user profile tags and channel characteristics, resulting in personalized marketing copywriting for each user.

[0049] Retrieve user profile tags for each user from the user profile database. User profile tags are structured information describing multiple dimensions of user characteristics, including demographic features (such as age, gender, and region), consumption preferences (such as beauty, digital products, and sports), historical purchase categories (such as the types of goods purchased in the past 30 days), active applications (the names of the applications most frequently used by the user), and behavioral tags (such as price sensitivity and preference for new products). Simultaneously, obtain the channel characteristics corresponding to the reach path selected for this user. Channel characteristics refer to the inherent constraints of different sending channels in terms of format and style. For example, application push channels support rich text and emojis, suitable for a friendly and lively tone. SMS service channels have strict character limits (usually within 70 Chinese characters) and do not support rich text, suitable for a concise, direct, and information-dense expression. In-app messaging channels support longer text and image-integrated content, suitable for explaining activity rules in detail.

[0050] The user profile tags and corresponding channel characteristics of each user are used as input parameters to call the copy generation interface of the natural language processing model. This natural language processing model adopts a generative pre-trained model based on a transformer architecture and has been fine-tuned on a large amount of historical marketing copy data and corresponding user feedback (click-through rate, conversion rate).

[0051] The model's processing logic is divided into two stages: content adaptation and format adaptation. In the content adaptation stage, the model identifies the most relevant points of interest from user profile tags. For example, if the user profile tags include fields such as "beauty preference: high" and "recently viewed: foundation", the model will determine that the core theme of the copy should revolve around foundation; if the user profile tags include "price sensitivity: high", the model will emphasize the discount in the copy.

[0052] During the format adaptation phase, the model adjusts the length, tone, and structure of the copy based on the characteristics of the input channel. For example, for app push notifications, the model generates long copy including emojis and friendly greetings. For SMS services, the model removes unnecessary modifiers, retains core information, and keeps the character count within limits. After this processing, the model outputs one or more candidate copy pieces, selecting the highest-rated one as the user's personalized marketing copy. For example, for a user whose profile tags include "beauty preference" and "female aged 20-25," if the reach path is app push notifications, the model might generate "Your favorite foundation is 20% off for a limited time! Click to claim your exclusive coupon." If the same user's reach path is SMS services, the model generates "Foundation 20% off, click to claim coupon." After generation, the system associates and stores each user's personalized marketing copy with their corresponding user identifier. Each copy integrates the user's individual characteristics and the dissemination characteristics of the selected channel.

[0053] In one embodiment, assembling the personalized marketing copy, the reach path, and the target correspondence, excluding the intercepted users, into a sending task, and sending the message corresponding to the sending task to the corresponding channel includes: Delete the entry corresponding to the intercepted user from the target mapping relationship to obtain a valid mapping relationship; The personalized marketing copy, the reach path, and the effective correspondence are assembled into a sending task, and the message corresponding to the sending task is sent to the corresponding channel.

[0054] Each entry in the target mapping relationship is associated with a user identifier and a target slot identifier. The blocked user list records the identifiers of users whose churn risk value exceeds the security threshold. The system iterates through each entry in the target mapping relationship, checking if the user identifier for that entry exists in the blocked user list. If it does, the entry is deleted. For example, if the original target mapping relationship includes users Zhang San, Li Si, and Wang Wu, and the blocked user list includes Zhang San, then deleting the entry corresponding to Zhang San will result in a valid mapping relationship that only includes Li Si and Wang Wu.

[0055] Personalized marketing copy, reach paths, and effective correspondence are assembled into a sending task. A sending task is a data package containing a task identifier, a list of users to be sent to, personalized marketing copy for each user, reach path for each user, and sending time for each user. The list of users to be sent to extracts user identifiers from the effective correspondence. Each user's personalized marketing copy is obtained from the previously generated copy set based on the user identifier. Each user's reach path is also obtained from the reach path results based on the user identifier. The sending time for each user is extracted from the time slot associated with each entry in the effective correspondence; for example, if the target slot is associated with 9 PM to 10 PM, then 9 PM is used as the sending time. After assembly, the sending task contains multiple marketing instructions to be executed, each instruction clearly specifying the time, channel, and user to which copy content should be sent.

[0056] The assembled sending task is sent to the message sending unit. Upon receiving the task, the message sending unit parses each instruction and sets a timer trigger based on the specified sending time. When the system time reaches that time, the message sending unit calls the corresponding channel gateway interface according to the reach path specified in the instruction, sending the personalized marketing copy to the target user's receiving endpoint. For example, if an instruction specifies that the reach path for user Li Si is application push, the sending time is 9 PM, and the personalized marketing copy is "Your favorite foundation is 20% off for a limited time," then the message sending unit will push this copy to Li Si's mobile device at 9 PM via the application push gateway. After the delivery is complete, the message sending unit will feed the result back to the data center for subsequent user profile updates and model training.

[0057] Reference Figure 2 The diagram shown is a functional module schematic of the message sending device 100 based on natural language processing according to this application.

[0058] Depending on the functions implemented, the message sending device 100 based on natural language processing includes a parsing module 110, a construction module 120, an evaluation module 130, a mapping module 140, a generation module 150, and a sending module 160. These modules can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, and are stored in the memory of the electronic device.

[0059] In this embodiment, the functions of each module / unit are as follows: Parsing module 110: Used to parse unstructured text instructions input by operators based on a natural language processing model to obtain structured basic strategy parameters; Construction module 120: used to construct the user's identity relationship graph with multiple channels, and aggregate the user's push frequency and reach load within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency and reach load; Evaluation module 130: is used to evaluate the expected benefits of multiple channel modes based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency and the reach load, in order to select the reach path and identify the blocked users; Mapping module 140: used to divide the sending window into multiple dynamic execution slots, map users to target slots according to users' historical active time preferences, and obtain the target correspondence between users and target slots; Generation module 150: Used to generate personalized marketing copy based on user profile tags and channel characteristics corresponding to the reach path through a natural language processing model; Sending module 160: Used to assemble the personalized marketing copy, the reach path and the target correspondence, excluding the intercepted users, into a sending task, and send the message corresponding to the sending task to the corresponding channel.

[0060] The specific implementation of the message sending device based on natural language processing in this application is largely the same as the specific implementation of the message sending method based on natural language processing described above, and will not be repeated here.

[0061] Reference Figure 3 The diagram shown is a schematic representation of a preferred embodiment of the electronic device of this application.

[0062] The electronic device includes a processor 111, a communication interface 112, a memory 113, and a communication bus 114, wherein the processor 111, the communication interface 112, and the memory 113 communicate with each other through the communication bus 114. The memory 113 is used to store computer programs, such as a message sending program based on natural language processing; Figure 3 Only an electronic device having a processor 111, a communication interface 112, a memory 113 and a communication bus 114 is shown. However, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0063] In one embodiment of this application, when the processor 111 executes a program stored in the memory 113, it implements the message sending method based on natural language processing provided in any of the foregoing method embodiments, including: The unstructured text instructions input by operators are parsed using a natural language processing model to obtain structured basic strategy parameters; Construct a user identity relationship graph with multiple channels, and aggregate the push frequency and reach load of users within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load, the expected benefits of multiple channel modes are evaluated to select the reach path and identify the blocked users. The sending window is divided into multiple dynamic execution slots. Users are mapped to target slots based on their historical active time preferences, thus obtaining the target correspondence between users and target slots. Based on user profile tags and the channel characteristics corresponding to the reach path, personalized marketing copy is generated through a natural language processing model. The personalized marketing copy, the reach path, and the target correspondence are combined into a sending task after excluding the intercepted users, and the message corresponding to the sending task is sent to the corresponding channel.

[0064] For a detailed description of the above steps, please refer to the flowchart of the above embodiment of the message sending method based on natural language processing.

[0065] Furthermore, this application also proposes a computer-readable storage medium that is both non-volatile and volatile. This computer-readable storage medium is any one or any combination of several of the following: hard disk, multimedia card, SD card, flash memory card, SMC, read-only memory (ROM), erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB memory, etc. The computer-readable storage medium includes a data storage area and a program storage area. The program storage area stores a message sending program based on natural language processing. When executed by a processor, the message sending program based on natural language processing performs the following operations: The unstructured text instructions input by operators are parsed using a natural language processing model to obtain structured basic strategy parameters; Construct a user identity relationship graph with multiple channels, and aggregate the push frequency and reach load of users within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load, the expected benefits of multiple channel modes are evaluated to select the reach path and identify the blocked users. The sending window is divided into multiple dynamic execution slots. Users are mapped to target slots based on their historical active time preferences, thus obtaining the target correspondence between users and target slots. Based on user profile tags and the channel characteristics corresponding to the reach path, personalized marketing copy is generated through a natural language processing model. The personalized marketing copy, the reach path, and the target correspondence are combined into a sending task after excluding the intercepted users, and the message corresponding to the sending task is sent to the corresponding channel.

[0066] The specific implementation of the computer-readable storage medium in this application is largely the same as the specific implementation of the message sending method based on natural language processing described above, and will not be repeated here.

[0067] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware simulation platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0069] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A message sending method based on natural language processing, characterized in that, The method includes: The unstructured text instructions input by operators are parsed using a natural language processing model to obtain structured basic strategy parameters; Construct a user identity relationship graph with multiple channels, and aggregate the push frequency and reach load of users within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load, the expected benefits of multiple channel modes are evaluated to select the reach path and identify the blocked users. The sending window is divided into multiple dynamic execution slots. Users are mapped to target slots based on their historical active time preferences, thus obtaining the target correspondence between users and target slots. Based on user profile tags and the channel characteristics corresponding to the reach path, personalized marketing copy is generated through a natural language processing model. The personalized marketing copy, the reach path, and the target correspondence are combined into a sending task after excluding the intercepted users, and the message corresponding to the sending task is sent to the corresponding channel. 2.The natural language processing based message sending method of claim 1, wherein, The method of parsing unstructured text instructions input by operators based on a natural language processing model yields structured basic strategy parameters, including: Input unstructured text instructions from operations personnel into a natural language processing model; The unstructured text instructions are segmented and intent is identified based on the natural language processing model to obtain segmentation results and intent identification results. Based on the word segmentation results and the intent recognition results, marketing objectives, cost preferences, and channel constraints are extracted from the unstructured text instructions; The marketing objectives, cost preferences, and channel constraints are combined into structured basic strategy parameters. 3.The natural language processing based message sending method of claim 1, wherein, The process of constructing a user identity relationship graph across multiple channels and aggregating the user's push frequency and reach load within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load includes: Obtain the user's identity identifier in various applications and SMS services, and establish the identity association relationship between the user and multiple channels; Based on the user's identity association with multiple channels, the push frequency and reach load of the user in each business line within the application matrix are statistically analyzed to obtain the user's push frequency and reach load in each business line. The user's identity association with multiple channels is constructed into a multi-channel identity relationship graph; The push frequency and reach load of the user in each business line are aggregated into the push frequency and reach load of the user in the application matrix, thus obtaining the user's multi-channel identity relationship graph, push frequency and reach load. 4.The natural language processing based message sending method of claim 1, wherein, The process of evaluating the expected returns of multiple channel modes based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load to select the reach path and identify blocked users includes: The structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load are input into a multi-objective machine learning model to obtain the expected conversion probability and expected delivery cost for each user. Based on each user's expected conversion probability and expected delivery cost, calculate each user's expected revenue under multiple channel models, and select the reach path with the highest expected revenue for each user to obtain each user's reach path; churn risk value for each user is calculated based on the multi-channel identity relationship graph, the push frequency, and the reach load. Users whose churn risk value exceeds a preset security threshold are marked as blocked users. 5.The natural language processing based message sending method of claim 1, wherein, The process of dividing the sending window into multiple dynamic execution slots and mapping users to target slots based on their historical active time preferences, thereby obtaining the target correspondence between users and target slots, includes: The sending window is divided into multiple dynamic execution slots based on hourly granularity and the number of channel modes; Obtain each user's historical active time preference, and map each user to a target slot among the multiple dynamic execution slots based on the user's historical active time preference, thereby obtaining the target correspondence between the user and the target slot. 6.The natural language processing based message sending method of claim 1, wherein, The process of generating personalized marketing copy based on user profile tags and channel characteristics corresponding to the reach path using a natural language processing model includes: Obtain user profile tags for each user and channel characteristics corresponding to the selected reach path for each user; Each user's user profile tags and channel characteristics are input into a natural language processing model. The natural language processing model then generates copywriting content that matches the user profile tags and channel characteristics, resulting in personalized marketing copywriting for each user. 7.The natural language processing based message sending method of claim 1, wherein, The step of assembling the personalized marketing copy, the reach path, and the target correspondence, excluding the intercepted users, into a sending task, and sending the message corresponding to the sending task to the corresponding channel includes: Delete the entry corresponding to the intercepted user from the target mapping relationship to obtain a valid mapping relationship; The personalized marketing copy, the reach path, and the effective correspondence are assembled into a sending task, and the message corresponding to the sending task is sent to the corresponding channel.

8. A message sending apparatus based on natural language processing, characterized by, The device includes: Parsing module: Used to parse unstructured text instructions input by operators based on natural language processing models to obtain structured basic strategy parameters; Construction module: Used to construct the user's identity relationship graph with multiple channels, and aggregate the user's push frequency and reach load within the application matrix to obtain the user's multi-channel identity relationship graph, push frequency, and reach load; Evaluation module: used to evaluate the expected benefits of multiple channel modes based on the structured basic strategy parameters, the multi-channel identity relationship graph, the push frequency, and the reach load in order to select the reach path and identify the blocked users; Mapping module: used to divide the sending window into multiple dynamic execution slots, and map users to target slots according to users' historical active time preferences, so as to obtain the target correspondence between users and target slots; Generation module: used to generate personalized marketing copy based on user profile tags and the channel characteristics corresponding to the reach path through a natural language processing model; Sending module: Used to assemble the personalized marketing copy, the reach path and the target correspondence, excluding the blocked users, into a sending task, and send the message corresponding to the sending task to the corresponding channel.

9. An electronic device, comprising: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the message sending method based on natural language processing as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the message sending method based on natural language processing as described in any one of claims 1 to 7.

Images