Intent recognition methods, computing devices, storage media and program products

Abstract

This application provides an intent recognition method, computing device, storage medium, and program product. The method involves detecting target user behavior that triggers a target condition, determining the target user corresponding to the target user behavior in response to the behavior, and extracting user behavior features from at least one set of behavior data corresponding to the target user. The at least one set of behavior data includes target behavior data corresponding to the target user behavior. A first intent recognition model is used to identify the target user's intent based on the user behavior features. The technical solution provided by this application can improve the accuracy of user intent recognition, thereby improving the accuracy of system recommendations.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intent recognition method, computing device, storage medium, and program product. Background Technology

[0002] With the rapid development of internet technology, the scale of online users and data has grown exponentially. Some online systems offer products such as goods, content, or web pages for users to consume. User intent has become a key factor in online systems' decision-making regarding their service content. For example, e-commerce systems that offer goods can provide personalized recommendations based on user intent.

[0003] Currently, user profiles are commonly used to depict user intent. However, the construction of user profiles is mainly based on long-term historical user behavior data (such as purchases, browsing history, and spending levels over the past 30 days). Therefore, the user profiles constructed have a significant lag. Due to the lag in user profiles, user intent is often not accurately identified during periods of abnormal behavioral density, such as promotional activities, holidays, or trending events (e.g., periods of high fluctuation in user behavior). This leads to lower accuracy in system recommendations and misses crucial opportunities to improve conversion rates. Summary of the Invention

[0004] This application provides an intent recognition method, computing device, storage medium, and program product to solve the problem of low system recommendation accuracy caused by inaccurate user intent recognition in the prior art.

[0005] Firstly, this application provides an intent recognition method, including: Detect the target user behavior that triggers the target condition; In response to the target user's behavior, determine the target user corresponding to the target user's behavior; User behavior features are extracted from at least one set of behavioral data corresponding to the target user; the at least one set of behavioral data includes target behavioral data corresponding to the target user's behavior. Using a first intent recognition model, the user intent of the target user is identified based on the user behavior characteristics.

[0006] Secondly, this application provides a computing device, including a processing component and a storage component; the storage component stores a computing program; the computer program is invoked and executed by the processing component to implement the intent recognition method as described in the first aspect above.

[0007] Thirdly, this application provides a computer storage medium storing a computer program thereon, which, when executed by a processing component, implements the intent recognition method as described in the first aspect above.

[0008] Fourthly, this application provides a computer program product, including a computer program or instructions, which, when executed by a processing component, implements the intent recognition method as described in the first aspect above.

[0009] This embodiment detects target user behavior that hits a target trigger condition. In response to the target user behavior, it determines the target user corresponding to the target user behavior and extracts user behavior features from at least one set of behavior data corresponding to the target user. The at least one set of behavior data includes target behavior data corresponding to the target user behavior. Using a first intent recognition model, based on the user behavior features, it identifies the target user's intent. This embodiment is driven by target user behavior that hits a target trigger condition. It extracts user behavior features from the real-time behavior sequence corresponding to the target user behavior (i.e., at least one set of user behavior data containing target behavior data), which are then used by the first intent recognition model to identify the user's intent. The identified user intent is the real-time user intent corresponding to the target user behavior, thus solving the lag defect of using user profiles to characterize user intent. Furthermore, this embodiment supports semantic reasoning capabilities for cold starts. Even in cold start scenarios, it can identify the user's real-time intent based on the target behavior data corresponding to the target user behavior, utilizing the semantic understanding capability of the first intent recognition model. Therefore, the solution in this embodiment greatly improves the accuracy of intent recognition, thereby helping to improve the accuracy of recommendations made by the system based on the user intent.

[0010] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of an embodiment of the intent recognition method provided in this application is shown; Figure 2 A schematic diagram of the structure of the first intent recognition model provided in this application is shown; Figure 3 A schematic diagram of the structure of the behavioral graph provided in this application is shown; Figure 4 The diagram illustrates the architecture of the intent recognition process in a real-world application scenario of this application. Figure 5A schematic diagram of the structure of an embodiment of the intent recognition device provided in this application is shown; Figure 6 A schematic diagram of the structure of the computing device provided in this application is shown. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user behavior data, user profile information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or generative models) comply with relevant laws and standards.

[0014] It should be noted that the technical solutions in this application are applicable to virtual network environments, and the users described generally refer to "virtual users." Real users can register user accounts on the server through registration to obtain user identities in the network environment. The same user account can log in to the server through different types of client terminals, enabling the server to identify the same user.

[0015] As described in the background technology above, user intent has become a key reference factor for online systems in deciding on their service content. Currently, user profiles are commonly used to characterize user intent. However, user profiles are mainly determined by statistical analysis of a large number of users' long-term historical behavioral data, resulting in stable long-term preferences of a group. This leads to a certain degree of lag and group-based bias. Therefore, in practical application environments (such as e-commerce environments that provide goods for consumption), the drawbacks of characterizing user intent based on user profiles are becoming increasingly apparent, especially during periods of abnormal behavioral density such as promotional activities, holidays, or trending events (e.g., periods of high fluctuation in user behavior). Because it is impossible to accurately grasp the real-time intent of individual users, the system may miss opportunities for recommendations, or the recommended content may be out of sync with the user's real-time intent, thus missing crucial opportunities to improve conversion rates.

[0016] To address this issue, the inventors conceived of introducing real-time stream computing technology to infer user intent based on real-time behavior sequences. This involves extracting behavioral features from the user's real-time behavior sequences and determining the user's intent through clustering and statistical methods. However, this approach emphasizes statistical patterns and heavily relies on historical behavior data, resulting in weak semantic understanding and poor performance in cold start scenarios.

[0017] To address this, the inventors conducted further research and proposed a solution for this application. The basic idea is to detect target user behavior that triggers the target condition, determine the target user corresponding to the target user behavior in response to the target user behavior, and extract user behavior features from at least one set of behavior data corresponding to the target user. The at least one set of behavior data includes target behavior data corresponding to the target user behavior. Using a first intent recognition model, the user intent of the target user is identified based on the user behavior features. This embodiment is driven by target user behavior that triggers the target condition, extracting user behavior features from the real-time behavior sequence corresponding to the target user behavior (i.e., at least one set of user behavior data containing target behavior data), which is then used by the first intent recognition model to identify the user intent. The identified user intent is the real-time user intent corresponding to the target user behavior, thus solving the lag defect in characterizing user intent using user profiles. Furthermore, this embodiment supports semantic reasoning capabilities for cold starts. Even in cold start scenarios, the real-time user intent can be identified based on the target behavior data corresponding to the target user behavior, utilizing the semantic understanding capability of the first intent recognition model. Therefore, the solution of this embodiment greatly improves the accuracy of intent recognition, thereby helping to improve the accuracy of system recommendations based on the user intent.

[0018] This embodiment realizes the transformation of user intent from "group intent" to "individual intent," from "offline intent" to "real-time intent," and from "intent determined by statistical patterns" to "intent recognized through semantic understanding." Through the natural language understanding capabilities of the first intent recognition model, user intent is grasped more accurately, helping the system provide timely and accurate decision support.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The implementation details of the technical solutions in the embodiments of this application are described in detail below.

[0021] Figure 1This is a flowchart of an embodiment of an intent recognition method provided in this application. The technical solution of this embodiment can be executed by a processing terminal, which can be a server in an online system (i.e., the target system described below), or it can be other nodes independent of the server in the online system.

[0022] In practical applications, online systems typically consist of a user terminal and a server terminal, which are connected via a network. The network provides the medium for communication links between the user terminal and the server terminal. Networks can include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0023] The client-side can be user-facing (i.e., system consumer), allowing users to perform interactive behaviors such as object searching and browsing. The client-side can interact with the server over the network to receive or send messages. For example, the client-side can sense user interactions and send corresponding interaction requests to the server; the server can process these requests and provide feedback to the client.

[0024] The user end can be a browser, an app (application), a web application such as an H5 (HyperText Markup Language 5) application, a mini-program (also known as a lightweight application), or a cloud application. The user end can be deployed on electronic devices and depends on the device to run or on certain apps within the device. Electronic devices can have displays and support information browsing, such as personal mobile terminals like smartphones, tablets, personal computers, desktop computers, smart speakers, smartwatches, etc.

[0025] The aforementioned processing or server may include servers that provide various services, such as servers that identify user intent based on user behavior triggered on the user's client, or servers that process interactive requests sent by the user's client.

[0026] It should be noted that the processing or server end can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. The server can also be a server in a distributed system, or a server combined with blockchain. The server can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0027] Figure 1The intent recognition method shown may include the following steps: S101, detect the target user behavior that hits the target trigger condition.

[0028] The target trigger condition can be a pre-defined rule used to determine whether user intent recognition should be triggered. For example, at least one user behavior that needs to trigger user intent recognition can be pre-determined, and specific trigger rules (such as behavior type, trigger frequency, context behavior, etc.) can be defined for each user behavior to form the target trigger condition. The target user behavior can be the user behavior of the target user that matches the target trigger condition.

[0029] This embodiment can scan the user behavior stream one by one through a real-time triggering mechanism (i.e., real-time trigger). When any user generates a user behavior, the user behavior is matched with a preset trigger condition. If the match is successful (i.e., a hit), the user behavior is determined to belong to the target user behavior, and the subsequent operations S102-S104 are initiated. If the match is unsuccessful (i.e., a miss), the process continues to wait for the detection of the next user behavior. This application completes the matching of the target trigger condition at the moment the user behavior occurs through a real-time triggering mechanism, thereby instantly initiating subsequent intent recognition operations based on the target user behavior, achieving millisecond-level response.

[0030] In practical applications, the user intent and response strategy reflected by the same user behavior may be completely different when triggered in different scenarios. Therefore, this embodiment can set multiple target scenarios and configure the target triggering conditions that match each target scenario. Different target scenarios correspond to different target triggering conditions to ensure that target user behaviors that truly have intent recognition needs are accurately selected for different scenario requirements.

[0031] In this embodiment, the "scenario" can refer to the environmental context in which the user's behavior occurs, and the "target scenario" is at least one pre-defined scenario with an intent recognition requirement. For example, it may include, but is not limited to: a search experience scenario (i.e., a scenario where the user performs a search operation on the target system), a third-party access scenario (i.e., a scenario where the user jumps from a third-party system to the target system), and a new user registration scenario (i.e., a scenario where the user registers an account on the target system). When the target scenario is a search experience scenario, the corresponding target triggering condition could be that the user performs a search operation, clicks on or browses any search result, etc. When the target scenario is a third-party access scenario, the corresponding target triggering condition could include that the user enters the target system through a Uniform Resource Locator (URL) carrying specific channel parameters, etc. When the target scenario is a new user registration scenario, the corresponding target triggering condition could be that the user enters the registration page and begins the registration process, or that the user has successfully completed registration, etc.

[0032] Optionally, this embodiment can adopt a similar approach to that described in the above embodiments to set corresponding target triggering conditions for each target scenario. Accordingly, when executing operation S101, it can detect target user behaviors that match the target triggering conditions corresponding to the target scenario. For example, by scanning the user behavior stream one by one through a real-time triggering mechanism (i.e., a real-time trigger), when any user generates a user behavior, it is determined whether the user behavior matches the target triggering condition corresponding to any target scenario; if so, the user behavior is taken as the target user behavior.

[0033] In some embodiments, to prevent the same user from triggering the same target user behavior multiple times, i.e., to prevent repeated triggering of real-time intent recognition for the same user, this embodiment can further superimpose suppression strategies on top of the target triggering conditions. For example, it can detect target user behavior that meets the target triggering conditions and satisfies the behavior information extraction requirements; the satisfies the behavior information extraction requirements include: the target user triggers the target user behavior for the first time within a first preset time interval from the current time corresponding to the target user behavior; or, detecting the generated object corresponding to the target user behavior. The generated object corresponding to the target user behavior can refer to the object generated after the target user behavior is executed. For example, if the target user behavior is a search behavior, the generated object can be the search result corresponding to the search behavior. It should be noted that if there is a lag or loading failure during the execution of the target user behavior, it indicates that the generated object corresponding to the target user behavior has not been detected.

[0034] In other words, after detecting user behavior that hits the target trigger condition, this embodiment needs to further determine whether the user behavior meets the information extraction requirements. That is, whether the target user triggered the user behavior for the first time within the first preset time period corresponding to the current time (i.e., within the detection period corresponding to the current time), or whether the user behavior has been completed and the corresponding generated object has been obtained. If the target trigger condition is hit and any of the above information extraction requirements are met, then the user behavior is taken as the target user behavior, and subsequent operations are performed.

[0035] S102, in response to the target user behavior, determine the target user corresponding to the target user behavior.

[0036] The target user can be the user who triggers the target user behavior. In practical applications, target user behavior usually carries a user identifier, such as a user's registered account on the target system. This embodiment can respond to the target user behavior, determine its corresponding user identifier, and then locate the target user who performed the target user behavior based on the user identifier.

[0037] S103, extract user behavior features from at least one behavioral data corresponding to the target user.

[0038] The at least one behavioral data corresponding to the target user can be a real-time behavioral data sequence associated with the target user. In a cold start scenario, it includes target behavioral data corresponding to the target user's behavior (i.e., behavioral data corresponding to the target user's behavior at the current moment); in a non-cold start scenario, in addition to including target behavioral data corresponding to the target user's behavior, it also includes behavioral data generated by the target user before the current moment. The user behavior features can be behavior-related features extracted from the behavioral data, such as, but not limited to, at least one of the following: behavioral object (i.e., the operation object corresponding to the user behavior), behavioral type (the type of user behavior), and generated object (i.e., the object generated by executing the user behavior). For example, if the user behavior is clicking a search control, then the behavioral object is "search control", the behavioral type is "search type", and the generated object is "search results".

[0039] One implementation of this embodiment is to extract user behavior features from the behavior data corresponding to the user behaviors that the target user has at the current time and in historical time, that is, to extract user behavior features from the behavior data corresponding to the behavior of all users associated with the target user.

[0040] Another approach is to acquire at least one set of behavioral data corresponding to at least one user action performed by the target user within a second preset time interval corresponding to the current moment of the target user's action, and then extract user behavior features from the at least one set of behavioral data. In other words, user behavior features are extracted from behavioral data corresponding to user actions performed by the target user at the current moment and within a period close to the current moment (i.e., the second preset time interval). This method provides user behavior features with better timeliness, avoiding interference from the current real-time intent recognition of user actions from historical moments distant from the current moment.

[0041] In practical applications, if the above S101 is to detect the target user behavior that hits the target trigger condition corresponding to the target scene, then this step can be to extract user behavior features from at least one behavior data of the target user corresponding to the target scene.

[0042] Specifically, this can involve filtering at least one behavioral data point corresponding to the target scenario from multiple behavioral data points associated with the target user, and then extracting user behavioral features from the filtered at least one behavioral data point. For example, for any target scenario, corresponding data filtering rules can be pre-defined, and a relationship can be established between the target scenario identifier and the corresponding data filtering rules. For example, if the target scenario is a search scenario, its corresponding filtering rules could be to filter behavioral data corresponding to search behavior, clicks on search results, exposure, transactions, and other behaviors.

[0043] In this embodiment, when extracting user behavior features from at least one set of behavioral data (such as at least one set of behavioral data corresponding to a selected target scenario), the user behavior features can be extracted from the at least one set of behavioral data using a feature extraction model or matching algorithm. Alternatively, user behavior data can be pre-collected and mapped to a behavior graph used to represent behavioral features. In this case, this step can be to extract the user behavior features corresponding to at least one set of behavioral data from the behavior graph. The specific implementation methods for constructing the behavior graph and extracting user behavior features from the behavior graph will be described in detail in subsequent embodiments.

[0044] S104 utilizes the first intent recognition model to identify the target user's intent based on user behavior characteristics.

[0045] Optionally, in this embodiment, user behavior features can be used as observable signals to quantify and characterize user behavior, and input into the first intent recognition model to control the intent recognition model to identify the user intent of the target user. In this embodiment, the so-called user intent is used to characterize the user's real-time interest tendency. If the user behavior features are extracted based on at least one behavioral data corresponding to the target scenario, then the user intent can be the unmet need reflected by the target user's behavior in a specific target scenario, such as the intent to compare prices, the intent to combine orders, or the intent to use image search to find lower-priced alternatives.

[0046] It should be noted that the models involved in the technical solutions provided in this application, such as the first intent recognition model in this step, and the second intent recognition model and feature extraction model mentioned below, can be deep learning models with relatively large model parameter scales. This application does not limit the number of model parameters supported by the deep learning model used, with the goal of meeting actual needs. The deep learning model involved in this application can be a language model (LM) or a multimodal model (MM).

[0047] In one implementation, such as Figure 2As shown, the first intent recognition model in this embodiment may include, for example, an input layer (Input) 10, an encoder (Encoder) 20, a decoder (Decoder) 30, and an output layer (Output) 40. It may also include a self-attention layer and a feed-forward neural network, etc., and this application does not impose any limitations on this. The input layer 10 is used to receive input data, such as user behavior features. The encoder 20 is mainly used to convert the input data into a vector representation; this process can incorporate the semantic features of the input data. The decoder 30 is responsible for converting the intermediate representation generated by the encoder 20 into output data (such as real-time user intent). The output layer 40 is used to output data, such as layout structure information in JSON format. A self-attention layer is a mechanism that allows the model to focus on other positions in the sequence to better encode the current position information. The feed-forward neural network can perform nonlinear transformations on the output of the self-attention layer to enhance the model's expressive power. The various parts work together, enabling the model built upon them to perform well in various complex processing tasks, such as natural language processing, computer vision, speech recognition, machine translation, text summarization, and intelligent question answering.

[0048] The first intent recognition model in this embodiment can be used in conjunction with a prompt to recognize the user's intent. The prompt instruction can include user behavior characteristics and intent recognition instructions, and can also include role information, recognition requirements, thought chain information, world knowledge base, input / output requirements, background data and / or example data, etc. Therefore, a prompt instruction can be generated based on intent recognition instructions and user behavior characteristics, as well as at least one of role information, recognition requirements, thought chain information, world knowledge base, input / output requirements, background data and example data, etc. The content items included in the prompt instruction can be set according to the actual situation, etc., and this application does not limit them.

[0049] Among them, intent recognition instructions explicitly tell the model the task operation to be performed, such as "In an e-commerce scenario, combine the characteristics of user shopping behavior (i.e., user behavior characteristics) to deeply identify and structure the behavioral intent (i.e., user intent) during the purchase process," etc.; role information can instruct the model to play a specific role, such as changing its professionalism, for example, "You are a professional intent analysis expert"; recognition requirements are used to constrain or standardize the task operation performed by the model, and can specify some constraints, such as recognition rules, to instruct the model to analyze from preset dimensions (such as price sensitivity, purchase strategy, brand preference, etc.); thought chain information can be used to guide the model to reason step by step; the world knowledge base can be a database that provides knowledge support for the model's divergent reasoning ability. Example data can be used to provide learning samples to help the model understand the operation to be performed, etc.; background data helps the model understand the task scenario, target audience, or preconditions, etc.; input / output requirements are used to standardize the format, length, style, tone, etc. of input / output data, such as "Output user intent in XX format." For ease of understanding, this prompt instruction can be, for example,: #Character Information: You are a professional user behavior analysis expert, skilled at accurately identifying users' true intentions by using multi-dimensional data such as user behavior sequences, time characteristics, and interaction depth on e-commerce platforms. #Intent recognition command: Based on the user behavior characteristics provided by the input data, identify the user's true intent; #Input / Output Requirements: Please output only a standard XX format object, without any other explanatory text. # Input data: ..." It should be noted that this example of prompt instructions is for illustrative purposes only and is not limited to this application. The content items included in the prompt instructions can be set or dynamically updated according to actual needs.

[0050] In addition, prompt templates can be pre-set. These templates are then concatenated with input data (i.e., user behavior characteristics) or filled into the "Input Data" field of the prompt template to generate corresponding prompt instructions. The prompt template can include intent recognition instructions, as well as one or more of the following: role information, recognition requirements, thought chain information, world knowledge base, input / output requirements, background data, and example data. Of course, one or more items in the prompt template can also be dynamically configured and take effect in real time, such as recognition requirements.

[0051] This embodiment utilizes the powerful semantic understanding and intelligent reasoning capabilities of the first intent recognition model to improve the accuracy of user intent recognition, thereby improving the accuracy of decision-making services based on the user intent.

[0052] The input data of the first intent recognition model in this embodiment may include the aforementioned user behavior features; the output data may be intent recognition results in a preset format (such as JOSN), wherein the intent recognition results include at least the recognized user intent, and may also include, but are not limited to: the confidence level of the recognized user intent, the basis for intent recognition, and other information generated during intent analysis and reasoning, such as user preferences based on user behavior feature analysis, further thinking results on user preferences, and user behavior scenarios based on user behavior feature analysis.

[0053] This embodiment converts target user behavior into contextual behavioral features described in natural language. Using prompts, it guides the first intent recognition model to perform semantic analysis, enabling it not only to recognize behavioral operations but also to understand their purpose, thus achieving a leap from behavioral appearance to the essence of intent. Furthermore, this embodiment fully utilizes the first intent recognition model's built-in world knowledge base and common-sense reasoning capabilities. For example, the recognition requirements of prompts can define common-sense reasoning capabilities that combine the world knowledge base and behavioral features to perform logical associations across categories and scenarios. For instance, when a user browses camping equipment, the first intent recognition model can infer, based on common sense, that the user may intend to purchase mosquito repellent spray or a portable power bank, rather than being limited to purchasing similar tents or sleeping bags, thereby overcoming the limitations of behavioral co-occurrence and expanding the diversity of recommendations.

[0054] This embodiment detects target user behavior that hits the target trigger condition. In response to this target user behavior, it determines the target user corresponding to the behavior and extracts user behavior features from at least one set of behavior data corresponding to the target user. The at least one set of behavior data includes target behavior data corresponding to the target user behavior. Using a first intent recognition model, based on these user behavior features, it identifies the target user's intent. This embodiment is driven by target user behavior that hits the target trigger condition. It extracts user behavior features from the real-time behavior sequence corresponding to the target user behavior (i.e., at least one set of user behavior data containing the target behavior data), which are then used by the first intent recognition model to identify the user's intent. The identified user intent is the real-time user intent corresponding to the target user behavior, thus overcoming the lag in characterizing user intent using user profiles. Furthermore, this embodiment supports semantic reasoning capabilities for cold starts. Even in cold start scenarios, it can identify the user's real-time intent based on the target behavior data corresponding to the target user behavior, utilizing the semantic understanding capabilities of the first intent recognition model. Therefore, the solution in this embodiment greatly improves the accuracy of intent recognition, thereby helping to improve the accuracy of recommendations based on the user intent.

[0055] Based on the above embodiments, this embodiment can also update the behavior graph used to characterize the user's global behavior features through the following sub-steps, and then extract user behavior features from the updated behavior graph.

[0056] Sub-step 1: In response to any user action triggered by any user, collect the behavioral data corresponding to that user action.

[0057] Wherein, any user can be any user of the target system, including the target user described in the above embodiments; any user behavior can be any user behavior triggered by any user, so it also includes the target user behavior described in the above embodiments.

[0058] This embodiment can detect any user behavior triggered by any user in the target system in real time and collect the corresponding behavioral data. In other words, during the update of the behavior graph, behavioral data is collected for any user behavior triggered by any user using the target system.

[0059] In practical applications, in order to further improve the comprehensiveness and accuracy of user behavior data acquisition, this embodiment collects behavioral data corresponding to any user behavior, including at least one of the following: acquiring behavioral data corresponding to any user behavior triggered by any user in the target system; acquiring behavioral data corresponding to any user jumping from a third-party system to the target system; and acquiring service data corresponding to any user behavior triggered by any user in the target system.

[0060] The target system can be an online system that provides related services and has a need to identify user intent; for example, it could be an e-commerce system that provides goods for purchase. The third-party system can be an external platform or channel independent of the target system, typically acting as a promotional partner or traffic source for the target system; for example, it could be social media or advertising platforms. The third-party system can embed redirect links carrying parameters (such as channel identifiers, the target system's service identifiers, or activity identifiers) to guide users to the corresponding page on the target system.

[0061] Specifically, for any user behavior triggered within the target system, the corresponding behavioral data within that system can be acquired, such as page view data, control click data, and element exposure data. This allows for real-time collection of comprehensive user behavior data across the target system. Furthermore, service data related to this user behavior within the target system can also be obtained. This service data includes server-side log data and service status data. Server-side log data may include server-side tracking logs and error code information related to user behavior. Service status data may include the status data of services provided by the target system. For example, if the target system is an e-commerce system, the corresponding service status data may include order data, payment data, coupon data, and verification data. If any user behavior involves a redirect from a third-party system to the target system, then the redirection record information corresponding to the redirection behavior can be collected, such as the redirection link clicked by the user, the service material clicked, and the channel.

[0062] Sub-step 2: Identify multiple entities from the behavioral data corresponding to any user behavior, as well as the relationships between these entities, and update the behavioral graph based on the multiple entities and their relationships.

[0063] In this behavioral graph, entities are represented as nodes, and relationships are represented as edges. Nodes and edges represent user behavior paths and relationships. In practical applications, entities in user behavior can include, but are not limited to: devices, browsed pages, clicked controls, exposed elements (such as products), third-party links, server responses, and service status. Each node records its corresponding entity attributes. For example, for a device node, its corresponding entity attribute could be the user identifier corresponding to the device that generated the user behavior, such as the user's login account. For a browsed page node, its corresponding entity attributes could include page browsing duration, session information, page source path, and special page information (such as product identifiers in product detail pages). For a clicked control node, its corresponding entity attributes could include control click information generated during page browsing, such as the clicked control identifier and location. For an exposed element node, its corresponding entity attributes could include element exposure information generated during page browsing, such as the exposed element identifier and location. For a third-party link node, its corresponding entity attributes could be parameters carried by the link, promotional materials, and channels. For server-side response nodes, their corresponding entity attributes can include server-side request response logs, such as various error code information.

[0064] The relationships between different entities may include, but are not limited to: the browsing relationship between a device and a browsing page; the click relationship between a browsing page and a clickable control; the exposure relationship between a browsing page and an exposed element; the jump relationship between different pages; the off-site click relationship between a third-party link and a page in the target system; and the response relationship between a browsing page and a server request, etc.

[0065] In practical applications, this embodiment can identify behavior objects from behavior data corresponding to any user behavior, as well as generated objects obtained by executing user behavior and behavior types of user behavior; treat behavior objects and generated objects as entities, use behavior types as the relationship between behavior objects and generated objects, and use user identifiers corresponding to any user as entity attributes to update the behavior graph.

[0066] Specifically, this embodiment can identify the behavior object, generating object, and behavior type from the behavior data corresponding to any user behavior using an entity relationship recognition model or a preset recognition algorithm (such as keyword matching, regular expression matching, etc.). It then determines whether the current behavior graph contains the behavior object and generating object. If they do, a relationship edge is established between the behavior object and generating object based on the behavior type, and the user identifier of the corresponding user (i.e., the user who triggered the user behavior) is added to the entity attribute information of the behavior object and generating object. If the current behavior graph does not contain the behavior object and generating object, new nodes are created for the behavior object and generating object in the behavior graph, and a relationship edge is established between the two nodes based on the behavior type. The user identifier of the corresponding user is also added to the entity attribute information of the behavior object and generating object. This embodiment can annotate the detailed information of user behavior data using entity attributes and edge attributes. By identifying detailed user behavior features from the behavior graph, it assists the first intent recognition model in better parsing user behavior information, thereby improving the accuracy of user intent recognition.

[0067] For example, Figure 3 A schematic diagram of the behavior graph of this embodiment is shown. "Device," "Page," "Third-party Link," "Server Response," "Control [Click]," and "Element [Exposure]" are points in the behavior graph that represent behavior objects or generating objects. "External Click," "Browse," "Request," "Exposure," "Click," and "Jump" are edges in the behavior graph that represent the relationship between behavior objects and generating objects. For each edge, the point pointed to by the arrow corresponds to the generating object, and the point pointed to by the other end of the edge corresponds to the behavior object.

[0068] This embodiment can be either maintaining a behavior graph for each user in the target system to record the user's behavior, or maintaining a shared behavior graph for all users in the target system, and there is no limitation on which one is used.

[0069] It should be noted that sub-step 1-sub-step 2 updating the behavior graph and the process of identifying user intent in S101-S104 are two independently executed links. That is, every time any user triggers any user behavior, sub-step 1 can be executed based on the user behavior, and S101 can be used to determine whether the user behavior hits the target triggering condition.

[0070] Based on the above principle that any user behavior of any user corresponds to updating the behavior graph, when performing the above S103 operation in this embodiment, it may include obtaining multiple target entities associated with the target user and the relationships between the multiple target entities from the behavior graph; and constructing user behavior features based on the multiple target entities and the relationships between the multiple target entities.

[0071] Since the behavior graph records behavioral data of any user's actions, this embodiment can, based on the target user's user identifier, search for multiple target entities associated with the target user and the relationships between them in the behavior graph, thus constituting user behavior features. This improves the comprehensiveness and efficiency of user behavior feature extraction.

[0072] In practical applications, the behavior graph records information related to the target user's full behavioral data, which is complex and massive. Furthermore, the context length of the first intent recognition model is limited, making it difficult to accurately identify user intent based on the behavioral features extracted from the full behavioral data. Additionally, different target scenarios require different behavioral data for intent recognition. Therefore, to improve the accuracy of intent recognition, this embodiment may determine the information filtering requirements for the target user's behavior corresponding to the target scenario; generate a graph query request based on the information filtering requirements; and, in response to the graph query request, filter multiple target entities from the behavior graph that are associated with the target user and the target scenario and meet the information filtering requirements, as well as the relationships between these multiple target entities.

[0073] Specifically, this embodiment can take the scenario corresponding to the target trigger condition triggered by the target user's behavior as the target scenario, and find the corresponding information filtering requirements based on the scenario identifier of the target scenario. The information filtering requirements are filtering requirements for behavioral data set for the target scenario. Different target scenarios correspond to different information filtering requirements. For example, if the target scenario is a search scenario, its corresponding filtering rules could be filtering the user's input search terms, the user's behavior on the search results page (such as sorting and filtering), the products distributed to the user, and their conversion status, etc. Since this data is only focused on a few core pages, the target entities and relationships filtered through the filtering requirements greatly reduce the context length of the first intent recognition model input.

[0074] This embodiment can generate a graph query request executable on the underlying storage medium based on the information filtering requirements corresponding to the target scenario. Responding to this graph query request, it filters multiple target entities from the behavior graph that are associated with the target user and the target scenario and meet the information filtering requirements, along with the relationships between these multiple target entities. Optionally, this embodiment can store the behavior graph in a database providing a graph query engine service (iGraph). It can then utilize the Gremlin syntax of the iGraph database to flexibly traverse the behavior graph and, according to the filtering requirements, query multiple target entities that are associated with the target user and the target scenario and meet the information filtering requirements, along with the relationships between these multiple target entities.

[0075] In practical applications, considering that the target entities and their relationships selected from the behavior graph may not be understood by the first intent recognition model, this embodiment can include the following when constructing user behavior features based on multiple target entities and their relationships: assembling multiple target entities and their relationships according to the preset expression method corresponding to the first intent recognition model to obtain user behavior features.

[0076] The preset expression method can be an expression method that the first intent recognition model can understand, which may include information encoding format, information assembly order, and arrangement format. The information encoding format can be the way entity information and relation information are encoded; for example, it can be encoded as key-value pairs. It should be noted that the encoding methods for entities and relations can be the same or different, and the encoding methods for different entities and different relations can also be the same or different; there is no limitation on this. The information assembly order can be the arrangement order when assembling multiple entities and relations. The information arrangement format can be the overall arrangement format of multiple entities and relations, the number of rows, the number of entities and relations corresponding to each row, and the separators between pairs of entities or relations, etc.

[0077] Specifically, assembling multiple target entities and the relationships between them according to the aforementioned preset expression method includes: obtaining entity information corresponding to each of the multiple target entities, and relationship information between the multiple target entities; encoding the entity information corresponding to each of the multiple target entities according to the information encoding format to obtain multiple entity encoding results, and encoding the relationship information between the multiple target entities to obtain multiple relationship encoding results; assembling the multiple entity encoding results and multiple relationship encoding results according to the information assembly order and arrangement format to obtain user behavior features. For example, the path a user visits to a page can be concisely expressed as "Homepage -> XX Channel -> Details Page -> Shopping Cart -> Order Page -> Payment Success Page" to facilitate understanding by the first intent recognition model.

[0078] Optionally, in some embodiments, to further improve the accuracy of user intent recognition, this embodiment may also store user profile information of all users of the target system and service profile information of the service objects in a database storing behavioral graphs (such as a Graph database). The user profile information can be a structured tag description of the target user based on multi-dimensional historical offline behavioral data and / or registration data, which can be used to characterize the user's inherent identity attributes, long-term interests, and behavioral habits. Compared to the real-time user intent output by the first intent recognition model, the user profile information focuses more on the user's long-term, stable user characteristics. The service profile information can be extracted from the multi-dimensional information of the service objects provided by the online system to characterize their semantic attributes, service characteristics, and applicable user groups.

[0079] Accordingly, when using the first intent recognition model to identify the user intent of a target user, the process can involve: acquiring user profile information and / or service profile information associated with user behavior features; and using the first intent recognition model to identify the user intent of the target user based on the user behavior features, user profile information, and / or service profile information. Specifically, after extracting user behavior features, the user profile information of the target user corresponding to the user behavior features is extracted from the maintained user profile information, as well as the service profile information of the service object corresponding to the user behavior features. In practical applications, this embodiment can decide whether to acquire user profile information or service profile information based on the target scenario associated with the user behavior features or the behavior type corresponding to the user behavior. After acquiring the user profile information and / or service profile information, it is used together with the user behavior features as the basic data for intent recognition and decision-making (i.e., the input data of the model), input into the first intent recognition model, and controlled to identify the user intent in a similar manner to that described in the above embodiment.

[0080] In practical applications, some scenarios already have clearly defined intent recognition rules based on accumulated industry knowledge and experience. For example, in a price comparison scenario, the corresponding intent recognition rules could be comparing prices of similar products within a short period of time, or using image search to find lower-priced alternatives. For the first intent recognition model, the assistance of intent recognition rules can more efficiently and accurately identify user intent. Therefore, in this embodiment, when using the first intent recognition model to identify the target user's intent based on user behavior characteristics, the intent recognition rules corresponding to the target scenario can be determined from the rule configuration data; the first intent recognition model can then be used to identify the target user's first user intent based on user behavior characteristics and according to the intent recognition rules.

[0081] Specifically, in this embodiment, for scenarios with defined intent recognition rules, the scenario identifier and its corresponding intent recognition rule are associated and stored in the rule configuration data. Before using the first intent recognition model for user identification, the intent recognition rule corresponding to the scenario identifier can be searched from the rule configuration data based on the scenario identifier of the target scenario. If found, a first prompt instruction is generated based on the found intent recognition rule, the first recognition requirement, the first intent recognition instruction, the first thought chain information, the role information, the world knowledge base, and / or the first example data. The first prompt instruction and user behavior characteristics are then input into the first intent recognition model. The first intent recognition model, according to the first prompt instruction, uses a combination of natural language parsing and intent recognition rule matching to determine the first user intent corresponding to the user behavior characteristics. It should be noted that the first user intent identified at this time is standardized and can be directly used by the target system to perform personalized recommendation tasks. It is a real-time intent tag of online users.

[0082] However, some scenarios may not have clearly defined intent recognition rules. That is, the intent recognition rule corresponding to the scenario identifier of the target scenario cannot be found in the rule configuration data, or the first intent recognition model fails to determine a standard first user intent based on the intent recognition rule. This embodiment refers to both of these situations as intent recognition rule determination failure. In this case, this embodiment further includes: if the intent recognition rule determination fails, using the first intent recognition model to perform semantic understanding of the user behavior features and generate intent description information; and determining the second user intent based on the intent description information.

[0083] Specifically, if the intent recognition rule fails to be determined, a second prompt instruction can be generated based on the second recognition requirement, the second intent recognition instruction, the second thought chain information, the role information, the world knowledge base, and / or the second example data. The second prompt instruction and the user behavior characteristics are then input into the first intent recognition model. The first intent recognition model, according to the second prompt instruction, uses its powerful natural language understanding and divergent reasoning capabilities to produce a summary of the user intent in natural language, that is, the intent description information corresponding to the user behavior characteristics.

[0084] It should be noted that while the role information and world knowledge base of the first and second prompt instructions mentioned above may be the same, the recognition requirements, recognition instructions, thought chain information, and example data may differ. For example, the first intent recognition instruction might be used in an e-commerce scenario, targeting user behavior characteristics and employing a combination of intent recognition rule matching and natural language parsing to deeply identify and structurally output standardized user intent tags. Correspondingly, the aforementioned first recognition requirements, first thought chain information, and first example data are specific recognition requirements, thought chain information, and example data set for how to deeply identify and structurally output standardized user intent tags through a combination of intent recognition rule matching and natural language parsing.

[0085] The second intent recognition instruction might be, in an e-commerce scenario, a description of the user's intent that is deeply identified and structured using natural language parsing, combined with user behavior characteristics. Correspondingly, the aforementioned second recognition requirements, second thought chain information, and second example data are specific recognition requirements, thought chain information, and example data set for how to combine user behavior characteristics to deeply identify and structure the description of the user's intent using natural language parsing. In other words, the first intent recognition model in this embodiment, when the intent recognition rules are determined, combines intent recognition rule matching with natural language parsing to output a standard user intent label. When the intent recognition rules fail to be determined, the powerful natural language parsing capability of the first intent recognition model is utilized to output a description of the user's intent.

[0086] In this embodiment, the intent description information identified by the first intent recognition model can be used to determine the target user's second user intent. One implementation method is to extract intent keywords from the real-time intent description information corresponding to the target user's behavior and convert them into standardized intent tags that can be directly used by the target system to perform personalized recommendation tasks, serving as the second user intent. In this case, the second user intent belongs to the real-time intent tag of the online user. This method of extracting intent keywords and converting intent tags can be determined by an artificial intelligence model, or it can be determined based on feature extraction algorithms, clustering algorithms, and tag mapping rules; there is no limitation on this.

[0087] Another implementation method is to use the user intent description information to statistically analyze the offline intent of the target user's group as its second user intent. In this case, the second user intent belongs to the offline intent tag of the target user's group. Optionally, this method can be implemented through the following two sub-steps: Sub-step A: Determine the target group to which the target users belong.

[0088] Specifically, this embodiment can determine the target group to which the target user belongs through at least one of the following methods.

[0089] Method 1: Determine the target group to which the target user belongs based on the target user's user profile information. Since the target system pre-constructs user profile information to represent any user, this embodiment can group users with the same identity attributes and preferences as the target user into a target group based on the target user's user profile information, and the target user belongs to that target group.

[0090] Method 2: The users corresponding to the generated intent description information within the statistical period are considered as the target group; the target users belong to the target group. This method can involve determining the users corresponding to each intent description information output by the first intent recognition model within the statistical period (e.g., one day) to which the target user behavior belongs, and treating these users as a target group. Since the target user behavior belongs to this statistical period, the target users corresponding to the target user behavior belong to this target group.

[0091] Method 3: Cluster the generated intent description information within the statistical period, and based on the clustering results, group the users corresponding to the same category of intent description information into a group, and designate the group to which the target user belongs as the target group. This method can cluster the intent description information output by the first intent recognition model within the statistical period (e.g., one day) to which the target user's behavior belongs, obtaining at least one set of intent description information. Each set of intent description information corresponds to a group of users, and finally, the group to which the target user belongs is designated as the target group.

[0092] This embodiment allows for flexible selection of the desired method to determine the target user's target group based on actual needs.

[0093] Sub-step B: Based on multiple intent descriptions corresponding to multiple users in the target group, extract at least one intent tag as the group intent of the target group; the group intent is used as the second user intent of the target user. For example, the group intent can be used as the second user intent of the target user.

[0094] One implementation method is to divide the multiple intent description information corresponding to multiple users in the target group into at least one information group according to similarity, and extract the intent tags corresponding to each group from the intent description information corresponding to at least one information group as the group intent of the target group.

[0095] Another approach is to extract key intent information from multiple intent descriptions corresponding to multiple users in the target group, divide the extracted key intent information into at least one information group according to similarity, summarize the key intent information of each information group, and obtain the intent label of each information group as the group intent of the target group.

[0096] In practical applications, to balance accuracy and efficiency in determining group intent, this embodiment can also employ different methods based on the number of users in the target group. The specific implementation is as follows: When the number of users in the target group is less than a preset value, a feature extraction model is used to extract key features from multiple intent descriptions corresponding to multiple users in the target group. These key features are then divided into at least one feature group. A second intent recognition model is then used to generate intent labels corresponding to each of the at least one feature group, which serve as the group intent of the target group. In other words, when the number of users in the target group is small, the powerful natural language understanding and divergent reasoning capabilities of the artificial intelligence model (i.e., the feature extraction model and the second intent recognition model) can be directly utilized to perform key feature extraction and intent label extraction. The operation of dividing the key features into at least one feature group (i.e., key feature clustering) can be implemented by the artificial intelligence model, for example, by the second intent recognition model, or through a clustering algorithm.

[0097] When the number of users in the target group is greater than or equal to a preset value, word segmentation is performed on multiple intent descriptions corresponding to multiple users in the target group. Key information is extracted from the word segmentation results, and candidate tags matching the key information are determined. At least one intent tag is determined from the candidate tags as the group intent of the target group. In other words, when the number of users in the target group is large, the method described above may be difficult to accurately and efficiently determine the group intent due to the limitation of the amount of context data of the artificial intelligence model. Therefore, this embodiment can perform word segmentation on multiple intent descriptions corresponding to multiple users in the target group, extract key information from the word segmentation results, and then determine candidate tags matching the key information by mapping the key information with intent tags (i.e., ontology mapping) or knowledge base matching. Finally, machine learning prediction or clustering algorithms are used to determine the intent tag from the candidate tags as the group intent of the target group.

[0098] In practical applications, the user intent determined in this embodiment can be used to perform recommendation operations to target users. Specifically, it can be based on a first user intent to perform a first recommendation operation to target users, or based on a group intent to perform a second recommendation operation to any user in the target group.

[0099] In this embodiment, the first user intent is the first user intent identified using the first intent recognition model based on the user behavior characteristics and according to the intent recognition rules. That is, it is the real-time user intent tag corresponding to the target user's execution of the target user behavior; for example, it could be the purchase intent of a target product determined based on the user's search behavior. The first recommendation operation performed based on the first user intent in this embodiment focuses on recommendation operations performed in real-time based on user behavior. This could include: allocating service content corresponding to the first user intent based on the user's real-time user behavior (i.e., the target user behavior) (e.g., a service object or activity information associated with that service object). For example, in an e-commerce application scenario, if the target user's first user intent is determined to be the purchase intent of a target product based on their real-time search behavior, the first recommendation operation performed on the target user could be sending recommendation information for the target product, allocating purchase incentive resources for the target product (such as discount coupons or electronic red envelopes), etc.

[0100] The group intent in this embodiment is based on a large amount of intent description information within a statistical period, and is obtained by offline statistical analysis for the target group. For example, the group intent could be interest in winter outdoor gear. The second recommendation operation performed based on the group intent in this embodiment focuses on recommendation operations performed periodically or when there is a need for content recommendation. For example, it could be periodically triggering the allocation of service content that matches the group intent; or it could be that when there is a need to promote certain content, the group intent is used to select the group that meets the promotion need, and the content is pushed to each user in that group.

[0101] In some embodiments, performing a first recommendation operation on the target user based on the first user intent includes: when the target user's behavior is in a period of abnormal behavior density, determining first content to be pushed to the target user based on the first user intent, and recommending the first content to the target user; wherein, the period of abnormal behavior density may be a period corresponding to a high-fluctuation scenario such as a promotional activity, a holiday, or a hot topic event. This is a good time for the target system to perform a recommendation operation, so it can determine service content (i.e., first content) that meets the target user's needs based on the target user's real-time user intent (i.e., the first user intent), and recommend the first content to the target user. For example, when the target user's behavior is in a period of abnormal behavior density, the target user and their first user intent can be sent to the recommendation system, and the recommendation system can allocate first content matching the target user's current needs based on the first user intent.

[0102] It should be noted that if this embodiment extracts the corresponding second user intent in real time based on the intent description information of the target user, the second user intent is also a real-time intent, so the first recommendation operation can also be performed on the target user based on the second user intent in this scenario.

[0103] In some embodiments, performing a second recommendation operation to any user in a target group based on group intent includes: when there is a demand for second content, determining whether the second content is suitable for the target group based on the group intent; if suitable, recommending the second content to any user in the target group. Specifically, since a user's real-time intent may only be a short-term intent rather than a long-term stable intent, when the target system has a content recommendation demand, this embodiment prioritizes selecting recommended users who are suitable for the second content based on group intent, that is, determining whether the group intent of the target group is suitable for the second recommended content; if so, pushing the second content to any user in the target group.

[0104] This embodiment employs different recommendation methods for real-time user intent (i.e., first user intent) and offline user intent (i.e., group intent), thereby improving the accuracy of system recommendations and thus enhancing the system's service conversion rate.

[0105] In a practical application scenario, an intelligent agent with intent recognition and application functions can be deployed on the processing end. This intelligent agent is based on the first intent recognition model described in the above embodiments and may also include functional modules such as user behavior collection, behavior graph update, graph information filtering, real-time trigger detection, and intent tag characterization. Next, this application embodiment takes an e-commerce system providing goods consumption, recommending goods based on user intent, as an example, combined with... Figure 4 This embodiment describes the process of recommending products based on user behavior and user intent.

[0106] In this embodiment, the intelligent agent deployed on the processing end can collect real-time user behavior data, that is, collect behavior data corresponding to any user behavior (i.e., in-site information) triggered by any user in the e-commerce system, such as page browsing behavior data, control click behavior data, login and registration behavior data, etc.; collect behavior data corresponding to any user jumping from a third-party system to the e-commerce system (i.e., off-site behavior); as well as server-side log data (i.e., backend logs) and service data representing service status (such as order data, payment data, coupon data, and verification data) of the e-commerce system (S401), and update the user behavior graph of the e-commerce system based on the collected user behavior data (S402).

[0107] On the other hand, the intelligent agent will also formulate corresponding real-time triggering mechanisms (i.e., real-time triggers) based on the actual needs of different scenarios. If there is a target user behavior that hits the target scenario (such as new user registration scenario, third-party landing scenario, and search experience scenario), the intelligent agent will receive the corresponding trigger message. This trigger message contains a scenario identifier, a user identifier, and a timestamp (S403). At this time, the information filtering requirements of the target scenario can be determined based on the scenario identifier. Then, based on the user identifier and the information filtering requirements, the agent will search for entity information and relationship information that occurred within the time period corresponding to the timestamp in the behavior graph for the corresponding target scenario, and assemble them into user behavior features according to a preset expression method (S404). For example, in the search scenario, the agent can filter search behavior, exposure details of cards on the page, click details of cards, and product transaction details. In the third-party landing scenario, the agent can filter external click behavior, advertising material information, landing page behavior tracking, and session behavior. In the new user registration scenario, the agent can filter registration and login behavior, user center logs, full site page information, and new user coupon information.

[0108] Optionally, in this embodiment, the underlying storage medium for storing the user behavior graph can be an iGraph database, which also stores offline user profiles and product profiles. Entity and relationship information is extracted using iGraph's Gremlin syntax, as well as real-time queries of associated user profile and product profile information. iGraph's Gremlin syntax supports highly flexible information traversal, filtering, and querying operations, and can ensure reduced response time in high-concurrency request scenarios. Furthermore, the structure of the results returned by Gremlin syntax, i.e., a JSON structure, is beneficial for the first intent recognition model to understand.

[0109] Then, depending on whether there is a clear intent recognition rule for the target scenario corresponding to the target user behavior, the system assembles prompt words corresponding to the target scenario (for example, if there is a clear intent recognition rule, then the intent recognition rule is added to the prompt words), and triggers the invocation of the natural language parsing model (i.e., the first intent recognition model). Combining the assembled prompt instructions with user behavior features, user profile information, product profile information, etc., extracted and assembled from the behavior graph, the system performs user intent recognition (S405). It should be noted that since the focus of intent recognition may differ in different target scenarios, this embodiment can customize specific prompt instructions for different target scenarios to ensure the accuracy of user intent recognition in different scenarios.

[0110] In this embodiment, for target scenarios with clear intent recognition rules, the natural language parsing model will output the labeled intent expression, namely the real-time standardized intent of the target user (S406), based on a combination of intent recognition rule matching and natural language parsing. For example, price comparison, ordering together, seasonal needs, brand preference, holiday needs, difficulty in making choices, etc.

[0111] For target scenarios without explicit intent recognition rules, the natural language parsing model performs semantic understanding and divergent reasoning based on user behavior characteristics to generate individual intent summaries (i.e., intent description information). It then determines the target user's group. If the target group is small (e.g., the number of users is less than a preset value), the model uses the individual intent summaries to characterize group labels. This involves extracting keywords and performing word vectorization, clustering, and model sampling summarization (i.e., using a second intent recognition model to summarize corresponding intent labels based on key features) to characterize the target group's group labels. If the target group is large (e.g., the number of users is greater than or equal to a preset value), a clustering statistical algorithm is used to characterize group labels based on the individual intent summaries. This involves text preprocessing (i.e., word segmentation of the individual intent summaries), keyword extraction, ontology mapping and knowledge base matching, clustering, and manual calibration to characterize the target group's group labels (S407).

[0112] It should be noted that users whose user intent is characterized by S406 belong to the real-time intent group; users whose group intent is characterized by S407 belong to the offline intent group. The real-time intent group and its corresponding real-time user intent, as well as the offline intent group and its corresponding group intent, are sent to the recommendation system so that the recommendation system can perform the corresponding recommendation operations (S408).

[0113] For example, a recommendation system may include a marketing platform, a triggering platform, and a user targeting platform. The marketing platform can respond to the real-time intent of users within a real-time user group and send coupons or other benefits to users in real time. The reach platform is responsible for pushing messages and sending emails to users. For example, it can identify users who are about to churn from offline intent groups and send them push messages, or it can work with the marketing platform to send coupons or other benefits to users who are about to churn through push messages, thereby guiding users back to consumption. The user targeting platform provides flexible user filtering capabilities. It can work with the marketing platform and the triggering platform to select users from the real-time user group to send messages or distribute benefits in real-time recommendation scenarios; it can also select users from the offline user group to send messages or distribute benefits in non-real-time recommendation scenarios.

[0114] This embodiment provides an intelligent agent architecture for real-time user intent recognition based on a natural language parsing model (i.e., a first intent recognition model). It deeply integrates the natural language parsing model into the user intent tag generation chain, achieving a paradigm shift from statistical features to semantic intent. This overcomes the shortcomings of offline static group tags. Furthermore, this embodiment introduces a scenario-based real-time triggering mechanism and a behavior graph filtering mechanism. By using scenario-corresponding target triggering conditions and data filtering rules, it achieves targeted and accurate querying of behavior data, resolving the contradiction between the excessive size of the full-volume behavior data and the context length limitation of the natural language parsing model. This embodiment also introduces a dual-mode output mechanism, supporting both real-time standardized intent tags (corresponding to the first user intent) and offline group intent output modes, achieving a balance between real-time operation and long-term profile construction. In addition, this embodiment adopts a Gremlin-based dynamic behavior graph query system to achieve high-concurrency, low-latency extraction of user behavior features and profile data, supporting millisecond-level model calls.

[0115] Furthermore, it's important to emphasize that this embodiment addresses the issues of delayed tag updates and inability to capture shifts in user interests inherent in traditional offline group intent tagging. It employs an event-driven, real-time trigger mechanism. This mechanism initiates intent recognition calculations the instant a user performs a key action (or target user action), combining the user's overall behavior graph with semantic reasoning from a natural language processing model to achieve sub-second response times. This enables the system to promptly grasp users' immediate needs in highly volatile scenarios, avoiding missed conversion windows.

[0116] Secondly, addressing the drawbacks of offline group intent tags being generic and ignoring individual differences, this embodiment abandons group statistical logic and instead infers personalized intent based on each user's complete behavioral context. The natural language processing model can understand the differentiated meanings of the same behavior across different users, thus outputting refined intent tags tailored to each individual, significantly improving user experience and service accuracy. Furthermore, compared to recommendation schemes that rely on real-time behavioral sequences and vector representations and are prone to information cocoons, this embodiment fully utilizes the built-in world knowledge and common-sense reasoning capabilities of the natural language processing model to achieve logical associations across categories and scenarios. This overcomes the limitations of behavioral co-occurrence and expands the diversity of recommendations.

[0117] Furthermore, this solution transforms user behavior into a natural language description context and guides the natural language processing model to perform semantic analysis using prompts. This enables the system to not only identify "what the user clicked" but also understand "why the user clicked," achieving a leap from behavioral appearance to the essence of intent. Moreover, in cold start scenarios, whether for new users or new products, this embodiment can make reasonable intent inferences based solely on limited target user behavior in the current session (such as search terms, page paths, dwell time, etc.) combined with the prior knowledge of the natural language processing model. This significantly alleviates the cold start dilemma and improves new user conversion rates and new product exposure efficiency.

[0118] Finally, addressing the issue of sparse user behavior and incomplete intent understanding due to focusing only on local clicks, this solution constructs a comprehensive user behavior graph covering both on-site and off-site activities, front-end and back-end, and explicit and implicit behaviors. Through Gremlin graph queries, multi-dimensional context is flexibly extracted, providing more comprehensive and coherent user behavior features for natural language processing models. This supports more accurate and robust intent recognition.

[0119] In other words, this embodiment not only solves the shortcomings of existing intent recognition in terms of timeliness, individuality, semantic depth, and cold start, but also achieves an intelligent upgrade from "passive response behavior" to "active understanding of intent" through the generalization and reasoning capabilities of the natural semantic parsing model, providing high-value real-time decision support for downstream services such as e-commerce recommendation, precision marketing, and user operation.

[0120] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0121] It should be noted that some processes described in the above embodiments and accompanying drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear in this document, or they may be executed in parallel. The operation numbers, such as S401, S403, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should also be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0122] Figure 5 A schematic diagram of an intent recognition device provided for an exemplary embodiment of this application is shown. The device includes: Behavior detection module 501 is used to detect the behavior of the target user that hits the target trigger condition; User determination module 502 is used to determine the target user corresponding to the target user behavior in response to the target user behavior; Feature extraction module 503 is used to extract user behavior features from at least one behavior data corresponding to the target user; the at least one behavior data includes target behavior data corresponding to the target user's behavior; The intent recognition module 504 is used to identify the user intent of the target user based on the user behavior characteristics using the first intent recognition model.

[0123] In some embodiments, the apparatus further includes: The data acquisition module is used to collect behavioral data corresponding to any user behavior triggered by any user; the arbitrary user includes the target user; the arbitrary user behavior includes the target user behavior; The graph update module is used to identify multiple entities from the behavior data corresponding to any user behavior, as well as the relationships between the multiple entities, and update the behavior graph based on the multiple entities and the relationships between the multiple entities; wherein, the entities are nodes in the behavior graph, and the relationships are edges in the behavior graph; The feature extraction module 503 is specifically used to obtain multiple target entities associated with the target user and the relationship between the multiple target entities from the behavior graph; and to construct user behavior features based on the multiple target entities and the relationship between the multiple target entities.

[0124] In some embodiments, the graph update module is specifically used to identify behavior objects from the behavior data corresponding to any user behavior, as well as generated objects obtained by executing the user behavior and the behavior type of the user behavior; to update the behavior graph by treating the behavior objects and the generated objects as entities, treating the behavior type as the relationship between the behavior objects and the generated objects, and treating the user identifier corresponding to any user as an entity attribute.

[0125] In some embodiments, the apparatus further includes: The filtering requirement determination module is used to determine the information filtering requirements for the target scenario corresponding to the target user behavior; The request generation module is used to generate a map query request based on the information filtering requirements. The feature extraction module 503 is specifically used to respond to the graph query request by filtering multiple target entities and the relationships between the multiple target entities from the behavior graph that are associated with the target user and the target scene and meet the information filtering requirements.

[0126] In some embodiments, the feature extraction module 503 is further configured to assemble the plurality of target entities and the relationships between the plurality of target entities according to a preset expression method corresponding to the first intent recognition model, so as to obtain user behavior features.

[0127] Optionally, the preset expression method includes: information encoding format, information assembly order, and arrangement format; when the feature extraction module 503 assembles the multiple target entities and the relationships between the multiple target entities, it is specifically used to: obtain entity information corresponding to the multiple target entities respectively, and relationship information between the multiple target entities; encode the entity information corresponding to the multiple target entities according to the information encoding format to obtain multiple entity encoding results, and encode the relationship information between the multiple target entities to obtain multiple relationship encoding results; assemble the multiple entity encoding results and the multiple relationship encoding results according to the information assembly order and arrangement format to obtain user behavior features.

[0128] In some embodiments, the data acquisition module specifically performs at least one of the following: acquiring behavioral data corresponding to any user behavior triggered by any user in the target system; acquiring behavioral data corresponding to any user jumping from a third-party system to the target system; acquiring service data corresponding to any user behavior triggered by any user in the target system.

[0129] In some embodiments, the behavior detection module 501 is specifically used to detect target user behavior that hits the target triggering condition corresponding to the target scene; the feature extraction module 503 is specifically used to extract user behavior features from at least one behavior data of the target user corresponding to the target scene.

[0130] In some embodiments, the intent recognition module 504 is specifically used to determine the intent recognition rule corresponding to the target scenario from the rule configuration data; and to recognize the first user intent of the target user based on the user behavior characteristics and the intent recognition rule using the first intent recognition model.

[0131] The intent recognition module 504 is further configured to, if the intent recognition rule fails to be determined, use the first intent recognition model to perform semantic understanding on the user behavior features and generate intent description information; and determine the second user intent based on the intent description information.

[0132] In some embodiments, the intent recognition module 504 is further configured to determine the target group to which the target user belongs; extract at least one intent tag as the group intent of the target group based on multiple intent description information corresponding to multiple users in the target group; and use the group intent as the second user intent of the target user.

[0133] In some embodiments, the intent recognition module 504 is further configured to, when the number of users in the target group is less than a preset value, use a feature extraction model to extract key features from multiple intent descriptions corresponding to multiple users in the target group, divide the key features into at least one feature group, and use a second intent recognition model to generate intent tags corresponding to each feature group as the group intent of the target group based on the key features corresponding to at least one feature group; when the number of users in the target group is greater than or equal to the preset value, perform word segmentation on multiple intent descriptions corresponding to multiple users in the target group, extract key information from the word segmentation results, determine candidate tags matching the key information, and determine at least one intent tag from the candidate tags as the group intent of the target group.

[0134] In some embodiments, the intent recognition module 504 is further configured to perform at least one of the following to determine the target group to which the target user belongs: determining the target group to which the target user belongs based on the user profile information of the target user; taking the users corresponding to the intent description information generated within the statistical period as the target group; the target user belonging to the target group; or, clustering the intent description information generated within the statistical period, and dividing the users corresponding to the intent description information of the same category into a group based on the clustering results, and taking the group to which the target user belongs as the target group.

[0135] In some embodiments, the apparatus further includes: The first recommendation module is used to perform a first recommendation operation to the target user based on the first user intent; The second recommendation module is used to perform a second recommendation operation on any user in the target group based on the group's intent.

[0136] In some embodiments, the first recommendation module is specifically used to determine the first content to be pushed to the target user based on the first user intent when the target user's behavior is in a period of abnormal behavior density, and recommend the first content to the target user; the second recommendation module is specifically used to determine whether the second content is suitable for the target group based on the group intent when there is a demand for second content, and if it is suitable, recommend the second content to any user in the target group.

[0137] In some embodiments, the behavior detection module 501 is further specifically used to detect target user behavior that hits the target triggering condition and meets the behavior information extraction requirements; the meeting of the behavior information extraction requirements includes: within a first preset time period from the current time corresponding to the target user behavior, the target user triggers the target user behavior for the first time; or, detecting the generated object corresponding to the target user behavior.

[0138] In some embodiments, the feature extraction module 503 is specifically used to obtain at least one behavior data corresponding to at least one user behavior that the target user has performed within a second preset time period from the current time corresponding to the target user behavior, and to extract user behavior features from the at least one behavior data.

[0139] Figure 5 The intent recognition device can perform Figure 1 The implementation principle and technical effects of the intent recognition method described in the illustrated embodiments will not be repeated here. The specific methods by which each module and unit of the intent recognition device in the above embodiments performs its operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0140] Figure 6 This is a schematic diagram of the structure of one embodiment of a computing device provided in this application. Figure 6 As shown, in practice, the computing device may include a storage component 601 and a processing component 602.

[0141] Storage component 601 is used to store computer programs and can be configured to store various other data to support operation on a computing device. Examples of this data include instructions for any application or method used to operate on the computing device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0142] Processing component 602, coupled to storage component 601, is used to execute computer programs in storage component 601 for implementing, etc. Figure 1 The diagram illustrates the identification method.

[0143] Furthermore, such as Figure 6 As shown, the computing device may also include other components such as a communication component 603, a display component 604, a power supply component 605, and an audio component 606. Figure 6 The diagram only shows some components and does not mean that the device includes only these components. Figure 6 The components shown. Additionally... Figure 6The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the computing device. The computing device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT (Internet of Things) device, or a server-side device such as a conventional server, cloud server, or server array. If the computing device in this embodiment is implemented as a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 6 The components within the dashed box; if the computing device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., it may be omitted. Figure 6 The component within the dashed box.

[0144] The processing component described above includes one or more processors to execute computer instructions to complete all or part of the steps in the method described above. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the method described above.

[0145] The aforementioned storage components can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0146] The aforementioned communication component is configured to facilitate wired or wireless communication between the device housing the communication component and other devices. The device housing the communication component can access wireless networks based on communication standards, such as mobile communication networks, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0147] The aforementioned display components may include a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0148] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0149] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0150] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a processor, the processor is able to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. In addition, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, so that the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device can be implemented as a means to implement the corresponding functions in the above method embodiments.

[0151] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0152] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0153] Finally, it should be noted that the above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An intent recognition method, characterized in that, include: Detect the target user behavior that triggers the target condition; In response to the target user's behavior, determine the target user corresponding to the target user's behavior; Extract user behavior features from at least one set of behavioral data corresponding to the target user; The at least one behavioral data includes target behavioral data corresponding to the target user's behavior; Using a first intent recognition model, the user intent of the target user is identified based on the user behavior characteristics.

2. The method according to claim 1, characterized in that, Also includes: In response to any user action triggered by any user, collect behavioral data corresponding to the arbitrary user action; Identify multiple entities from the behavioral data corresponding to any user behavior, as well as the relationships between the multiple entities, and update the behavioral graph based on the multiple entities and the relationships between the multiple entities; wherein, the entities are nodes in the behavioral graph, and the relationships are edges in the behavioral graph; The step of extracting user behavior features from at least one behavioral data corresponding to the target user includes: From the behavior graph, obtain multiple target entities associated with the target user and the relationships between the multiple target entities; User behavior features are constructed based on the multiple target entities and the relationships between them.

3. The method according to claim 2, characterized in that, The step of identifying multiple entities from the behavioral data corresponding to any user behavior, as well as the relationships between the multiple entities, and updating the behavioral graph based on the multiple entities and the relationships between the multiple entities includes: Identify the behavior object from the behavior data corresponding to any user behavior, as well as the generated object obtained by executing the user behavior, and the behavior type of the user behavior; The behavior graph is updated by treating the behavior object and the generated object as entities, the behavior type as the relationship between the behavior object and the generated object, and the user identifier corresponding to any user as an entity attribute.

4. The method according to claim 2, characterized in that, Also includes: Determine the information filtering requirements for the target scenario corresponding to the target user behavior; Based on the aforementioned information filtering requirements, a map query request is generated; The step of obtaining multiple target entities associated with the target user from the behavior graph and the relationships between the multiple target entities includes: In response to the graph query request, multiple target entities and the relationships between the multiple target entities are filtered from the behavior graph that are associated with the target user and the target scenario and meet the information filtering requirements.

5. The method according to claim 2, characterized in that, The construction of user behavior features based on the multiple target entities and the relationships between them includes: According to the preset expression method corresponding to the first intent recognition model, the multiple target entities and the relationships between the multiple target entities are assembled to obtain user behavior features.

6. The method according to claim 5, characterized in that, The preset expression method includes: information encoding format, information assembly order, and arrangement format; the assembly of the multiple target entities and the relationships between the multiple target entities includes: Obtain entity information corresponding to each of the multiple target entities, as well as relationship information between the multiple target entities; According to the information encoding format, the entity information corresponding to multiple target entities is encoded to obtain multiple entity encoding results, and the relationship information between the multiple target entities is encoded to obtain multiple relationship encoding results; According to the information assembly order and arrangement format, the multiple entity encoding results and the multiple relation encoding results are assembled to obtain user behavior features.

7. The method according to claim 2, characterized in that, The collected behavioral data corresponding to any user behavior includes at least one of the following: Get behavioral data corresponding to any user behavior triggered by any user in the target system; Obtain behavioral data of any user who jumps from a third-party system to the target system; Obtain service data corresponding to any user behavior triggered by any user in the target system.

8. The method according to claim 1, characterized in that, The target user behaviors that trigger the detection conditions include: Detect the target user behavior that matches the target trigger condition in the target scenario; The step of extracting user behavior features from at least one behavioral data corresponding to the target user includes: User behavior features are extracted from at least one set of behavioral data of the target user corresponding to the target scenario.

9. The method according to claim 4 or 8, characterized in that, The step of using the first intent recognition model to identify the target user's intent based on the user behavior characteristics includes: Determine the intent recognition rules corresponding to the target scenario from the rule configuration data; Using the first intent recognition model, based on the user behavior characteristics, and in accordance with the intent recognition rules, the first user intent of the target user is identified.

10. The method according to claim 9, characterized in that, Also includes: If the intent recognition rule fails to be determined, the first intent recognition model is used to perform semantic understanding on the user behavior features and generate intent description information. Based on the intent description information, determine the second user intent.

11. The method according to claim 10, characterized in that, Determining the second user intent based on the intent description information includes: Determine the target group to which the target user belongs; Based on multiple intent descriptions corresponding to multiple users in the target group, at least one intent tag is extracted as the group intent of the target group; the group intent is used as the second user intent of the target user.

12. The method according to claim 11, characterized in that, The step of extracting at least one intent tag as the group intent of the target group based on multiple intent descriptions corresponding to multiple users in the target group includes: When the number of users in the target group is less than a preset value, a feature extraction model is used to extract key features from multiple intent descriptions corresponding to multiple users in the target group, and the key features are divided into at least one feature group. A second intent recognition model is used to generate intent labels corresponding to each feature group as the group intent of the target group based on the key features corresponding to at least one feature group. When the number of users in the target group is greater than or equal to the preset value, the multiple intent description information corresponding to multiple users in the target group is segmented into words, key information is extracted from the segmentation results, and candidate tags matching the key information are determined. At least one intent tag is determined from the candidate tags as the group intent of the target group.

13. The method according to claim 11, characterized in that, The method for determining the target group to which the target user belongs includes at least one of the following: Based on the user profile information of the target user, the target group to which the target user belongs is determined; The users corresponding to the intent description information generated within the statistical period are taken as the target group; The target users belong to the target group; or, The intent description information generated within the statistical period is clustered, and based on the clustering results, users corresponding to the same category of intent description information are divided into a group, and the group to which the target user belongs is taken as the target group.

14. The method according to claim 11 or 12, characterized in that, Also includes: Based on the first user intent, a first recommendation operation is performed on the target user; Based on the group's intent, a second recommendation operation is performed on any user within the target group.

15. The method according to claim 14, characterized in that, The step of performing the first recommendation operation to the target user based on the first user intent includes: When the target user's behavior is in a period of abnormal behavior density, the first content to be pushed to the target user is determined based on the first user intent, and the first content is recommended to the target user. The step of performing the second recommendation operation to any user in the target group based on the group intent includes: When there is a need to push second content, it is determined whether the second content is suitable for the target group based on the group's intent. If it is suitable, the second content is recommended to any user in the target group.

16. The method according to claim 1, characterized in that, The target user behaviors that trigger the detection conditions include: Detect target user behavior that meets the triggering conditions and satisfies the requirements for extracting behavioral information; The requirements for satisfying behavioral information extraction include: Within a first preset time interval from the current moment corresponding to the target user's behavior, the target user triggers the target user behavior for the first time; Alternatively, the generated object corresponding to the target user's behavior can be detected.

17. The method according to claim 1, characterized in that, The step of extracting user behavior features from at least one behavioral data corresponding to the target user includes: Within a second preset time interval corresponding to the current time of the target user's behavior, at least one behavioral data corresponding to at least one user behavior of the target user within the second preset time interval is obtained, and user behavior features are extracted from the at least one behavioral data.

18. A computing device, characterized in that, This includes processing components and storage components; The storage component stores a computer program; the computer program is invoked and executed by the processing component to implement the intent recognition method as described in any one of claims 1-17.

19. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processing component, implements the intent recognition method as described in any one of claims 1-17.

20. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processing component, implement the intent recognition method as described in any one of claims 1-17.

Images