An advertisement accurate delivery method, system and platform based on a customer resource library

Abstract

The application relates to the technical field of data processing, and particularly discloses an advertisement accurate putting method, system and platform based on a customer resource library, which comprises the following steps: in response to a login action triggered by a customer accessing a website, extracting attribute characteristic data of the customer from the customer resource library; in response to a real-time consultation event, obtaining a real-time consultation interaction data stream, performing knowledge extraction on the attribute characteristic data of the customer, and obtaining a basic customer knowledge graph; wherein the basic customer knowledge graph comprises semantic correlation between customer attributes and customer behaviors; performing knowledge fusion on the real-time consultation interaction data stream and the basic customer knowledge graph, and obtaining a dynamically updated customer intention knowledge graph; in response to the intention state change of a target customer in the customer intention knowledge graph satisfying a preset trigger condition, determining a target advertisement type according to a current intention label of the target customer, and calling the corresponding target advertisement to perform real-time putting.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, and platform for precise advertising targeting based on a customer resource database. Background Technology

[0002] With the continuous development of internet advertising technology, precise targeting has become a key means to improve advertising conversion rates. Existing precise advertising methods usually build customer profiles based on customers' historical behavior data, predict customers' interests and preferences through rule matching or machine learning models, and then push advertisements.

[0003] However, existing technologies have the following shortcomings: on the one hand, traditional methods only understand customers at the level of static labels, making it difficult to capture the dynamic changes in customers' real-time intentions; on the other hand, when customer data comes from multiple heterogeneous systems, there are semantic heterogeneity and entity ambiguity issues between data from different sources, making it difficult to form a unified and accurate understanding of customer intentions, thereby affecting the accuracy of advertising.

[0004] Therefore, this invention proposes a method, system, and platform for precise advertising delivery based on a customer resource database. Summary of the Invention

[0005] This invention provides a method, system, and platform for precise advertising delivery based on a customer resource database, in order to solve the problems of static customer intent understanding and difficulty in integrating multi-source heterogeneous data in the prior art, and to achieve real-time intent-driven precise advertising delivery.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for precise ad targeting based on a customer resource database is provided, including: In response to a customer's login behavior triggered by visiting the website, the system extracts the customer's attribute data from the customer database; the customer attribute data includes the customer's login information and browsing information. In response to real-time consultation events, acquire real-time consultation interaction data streams, which include at least real-time consultation text and / or transcribed text of consultation quotes; Knowledge is extracted from customer attribute feature data to obtain a basic customer knowledge graph; the basic customer knowledge graph includes the semantic relationship between customer attributes and customer behavior. By integrating real-time consultation interaction data streams with basic customer knowledge graphs, a dynamically updated customer intent knowledge graph is obtained. In response to changes in the target customer's intent state in the customer intent knowledge graph that meet preset trigger conditions, the target ad type is determined based on the target customer's current intent tag, and the corresponding target ad is retrieved and delivered in real time.

[0007] This method extracts knowledge from customer attribute feature data to obtain a basic customer knowledge graph. This graph contains semantic relationships between customer attributes and customer behaviors, achieving a structured expression of the static customer profile. Based on this, real-time consultation interaction data streams are fused with the basic customer knowledge graph to establish semantic relationships between real-time extracted intent entities and static customer entities, resulting in a dynamically updated customer intent knowledge graph. When a change in the target customer's intent state in the customer intent knowledge graph meets preset trigger conditions, the real-time delivery of the target advertisement is triggered. This method achieves a leap from static customer profiles to dynamic intent-driven advertising delivery, significantly improving the timeliness and accuracy of advertising delivery.

[0008] In one possible implementation, knowledge extraction is performed on customer attribute feature data to obtain a basic customer knowledge graph. This includes: data identification of customer attribute feature data to distinguish between structured and unstructured data; for structured data, customer attribute entities are extracted from it using preset mapping rules; for unstructured data, customer behavior entities are extracted from it using named entity recognition technology, and semantic relationships between customer attribute entities and customer behavior entities are identified using relation extraction technology; the extracted customer attribute entities, customer behavior entities, and semantic relationships are organized into preliminary knowledge units in the form of triples; and based on a customer domain ontology model, knowledge fusion is performed on the preliminary knowledge units to eliminate entity ambiguity from different data sources and integrate semantic relationships to generate integrated knowledge units. Based on the customer domain ontology model, the entity types and relationship types of the basic customer knowledge graph are defined, and the integrated knowledge units are used as instance data to fill in the corresponding data to construct the basic customer knowledge graph.

[0009] In one possible implementation, real-time consultation interaction data streams are fused with a basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph. This includes: performing streaming knowledge extraction on the real-time consultation interaction data streams to extract intent entities, consultation focuses, and contextual semantic relationships between intent entities and consultation focuses from real-time consultation text and / or transcribed text of inquiry quotes, forming real-time intent knowledge units; and fusing the real-time intent knowledge units with the basic customer knowledge graph based on a cross-domain intent ontology model to obtain a dynamically updated customer intent knowledge graph. The cross-domain intent ontology model is used to define the semantic hierarchy and relational constraints between preset knowledge categories, supporting the establishment of entity links and semantic relationships between intent entities in the real-time intent knowledge units and customer entities in the basic customer knowledge graph.

[0010] In one possible implementation, based on a cross-domain intent ontology model, real-time intent knowledge units are fused with a basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph. This includes: linking intent entities in the real-time intent knowledge units with customer entities in the basic customer knowledge graph to identify customer entities that point to the same real-world entity as the intent entities; customer entities are nodes representing customers in the basic customer knowledge graph; intent entities are nodes representing customer consultation intent in the real-time consultation interaction data stream; the real-world entity is the customer that generates the real-time consultation interaction data stream; based on the relational constraints defined by the cross-domain intent ontology model, real-time semantic associations are established between intent entities and identified customer entities, with the type of real-time semantic association determined by the relational constraints; the consultation focus and contextual semantic relationships associated with the intent entities are added to the customer intent knowledge graph as attribute information of the real-time semantic associations to obtain a dynamically updated customer intent knowledge graph.

[0011] In one possible implementation, the intent entities in the real-time intent knowledge unit are linked with the customer entities in the basic customer knowledge graph. This includes: extracting the attribute features of the intent entities and the attribute features of the customer entities; calculating the attribute similarity between the intent entities and each customer entity; and identifying the customer entity with the highest similarity exceeding a preset threshold as the customer entity that points to the same real-world subject as the intent entity.

[0012] In one possible implementation, in response to a change in the target customer's intent state in the customer intent knowledge graph satisfying a preset trigger condition, the target ad type is determined based on the target customer's current intent tag, and the corresponding target ad is retrieved and delivered in real time. This includes: monitoring the state changes of the intent nodes corresponding to the target customer in the customer intent knowledge graph, where the state changes are triggered by knowledge processing operations performed on the customer intent knowledge graph, including intent reasoning based on ontology rules and / or intent relationship completion based on graph embedding; when the target customer's intent state change is detected to satisfy the preset trigger condition, the target customer's current intent tag is extracted; based on the preset mapping relationship between intent tags and ad types, the target ad type corresponding to the current intent tag is determined; and a target ad matching the target ad type is retrieved from the ad resource library and delivered to the target customer in real time.

[0013] In one possible implementation, the knowledge processing operation includes: acquiring the intent entities, consultation focus, and real-time semantic associations associated with the target customer in the customer intent knowledge graph; performing deductive reasoning on the attributes of the intent entities and real-time semantic associations based on preset ontology rules to update the target customer's intent state; and / or, using a knowledge graph embedding model to predict the missing intent entities or real-time semantic associations in the customer intent knowledge graph, completing the customer intent knowledge graph based on the prediction results, and updating the target customer's intent state based on the completed graph.

[0014] In one possible implementation, the preset triggering conditions include at least one of the following: the real-time consultation frequency of the target customer exceeds a preset threshold; the customer intent knowledge graph indicates that the risk intent level of the target customer is upgraded from level one to level two; level one is lower than level two; the target customer generates a quote inquiry behavior within a preset time window.

[0015] Secondly, a precise advertising delivery system based on a customer resource database is provided, including: The attribute data acquisition module is used to extract customer attribute feature data from the customer resource database in response to the login behavior triggered by a customer visiting the website; the customer attribute feature data includes the customer's login information and browsing information; The real-time data acquisition module is used to respond to real-time consultation events and acquire real-time consultation interaction data streams, which include at least real-time consultation text and / or transcribed text of consultation quotes. The knowledge extraction module is used to extract knowledge from customer attribute feature data to obtain a basic customer knowledge graph; the basic customer knowledge graph includes the semantic relationship between customer attributes and customer behavior. The knowledge fusion module is used to fuse real-time consultation interaction data streams with basic customer knowledge graphs to obtain dynamically updated customer intent knowledge graphs. The ad delivery module is used to respond to changes in the target customer's intent state in the customer intent knowledge graph that meet preset trigger conditions, determine the target ad type based on the target customer's current intent tag, and retrieve the corresponding target ad for real-time delivery.

[0016] Thirdly, it provides a precise advertising platform based on a customer resource database, including: The professional knowledge service subsystem is used to provide professional knowledge training and quotation services, generate real-time consultation events, and output real-time consultation interaction data streams; The customer growth management subsystem is used to provide customer growth management services and to transmit customer conversion behavior data back. And, as in the second aspect, a precise advertising delivery system; The targeted advertising system establishes data connections with the professional knowledge service subsystem and the customer growth management subsystem to receive real-time consultation interaction data streams and customer conversion behavior data.

[0017] The beneficial effects of this invention are as follows: By extracting knowledge from customer attribute and feature data and achieving knowledge fusion based on a customer domain ontology model, a basic customer knowledge graph containing semantic relationships is constructed. This is then integrated with real-time consultation and interaction data streams. Through knowledge processing, intent reasoning and relationship completion are achieved, resulting in a dynamically updated customer intent knowledge graph. With the full lifecycle technical support of knowledge graphs—including knowledge extraction, knowledge fusion, and knowledge processing—advertising is upgraded from static customer profiles to real-time intent-driven campaigns, significantly improving the accuracy and conversion rate of advertising. Simultaneously, the data linkage between the professional knowledge service platform and the customer growth management platform ensures that the effects of each campaign are fed back into the knowledge graph.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating a method for precise ad targeting based on a customer resource database, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the basic customer knowledge graph construction process provided in this embodiment of the invention; Figure 3 A flowchart illustrating the customer intent knowledge graph construction process provided in an embodiment of the present invention; Figure 4 A flowchart illustrating the knowledge fusion process provided in an embodiment of the present invention; Figure 5 A flowchart illustrating the advertising placement decision-making process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an advertising precision targeting system based on a customer resource database, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structure of an advertising precision targeting platform based on a customer resource database, provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] It should be noted that, in this embodiment of the invention, "customer" refers to the entity that interacts with the system, such as a natural person who visits a website or initiates an online consultation; "customer" refers to the commercial service object stored in the customer resource database. While both "customer" and "target customer" refer to the same real-world entity, they emphasize different roles in different data processing stages. In this context, both "target customer" and "target client" refer to the specific customer who triggers the advertising placement decision. Example 1

[0023] Figure 1 This is a flowchart illustrating a method for precise ad targeting based on a customer resource database, provided as an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step S101: In response to the login behavior triggered by the customer accessing the website, extract the customer's attribute feature data from the customer resource database.

[0024] In one possible implementation, when a customer accesses the website and completes login authentication via a terminal device, the system automatically triggers the extraction of attribute feature data.

[0025] The aforementioned customer resource database is a pre-built and maintained customer data storage system. Its data sources include, but are not limited to: basic information submitted by customers during website registration, session logs generated from each customer login, behavioral tracking data generated when customers browse website pages, and interaction records left by customers when participating in marketing activities. The customer resource database can be implemented using a relational database (such as MySQL or PostgreSQL) combined with a distributed file system (such as HDFS) to uniformly store and manage structured field data and semi-structured log data.

[0026] Customer attribute data extracted from the customer resource database specifically includes the following two categories: (1) Login information: including but not limited to customer login ID, login timestamp, login device type (e.g., PC, mobile, tablet), device operating system version, browser type and version, IP address and its location (down to the city level), login method (e.g., account and password login, mobile verification code login, third-party authorized login), etc. Login information helps the system confirm customer identity and establish a session tracking link.

[0027] (2) Browsing information: This includes, but is not limited to, the Uniform Resource Locator (URL) of the website pages visited by the customer, page titles and page type tags (such as product introduction pages, price comparison pages, claims guide pages, FAQ pages, etc.), page dwell time, page scrolling depth, the sequence of clicks within the page (such as which buttons, links, or tabs were clicked), keywords entered in the search box during browsing, and the browsing path (i.e., the sequence of pages visited by the customer from entering the website to leaving the website). Browsing information can reflect the customer's interests, preferences, and behavioral intentions.

[0028] It should be noted that the login ID in the login information is the core identifier associated with the customer's identity. The system will use this ID as a clue to link and integrate the attribute feature data and real-time consultation interaction data stream of the same customer in subsequent steps.

[0029] Step S102: In response to a real-time consultation event, obtain the real-time consultation interaction data stream.

[0030] Real-time consultation events are generated by professional knowledge service platforms, such as customers initiating insurance knowledge consultations or requesting quotes through online customer service systems.

[0031] Specifically, when a customer initiates a consultation request through the online customer service portal embedded on the website page, the professional knowledge service platform detects the establishment of a consultation session, generates a real-time consultation event, and pushes a notification to the system.

[0032] For example, when a real-time consultation event is triggered, the system obtains the real-time consultation interaction data stream corresponding to that customer.

[0033] The professional knowledge service platform is an independently deployed service system used to provide clients with professional knowledge consultation and quotation services. Taking the insurance industry as an example, the platform can be deployed in the Chengdu operations center to provide clients nationwide with professional knowledge sharing, online Q&A, and quotation services for non-motor insurance (such as property insurance, liability insurance, credit insurance, guarantee insurance, etc.). The platform supports multiple interaction methods, including online text chat and voice calls.

[0034] The aforementioned real-time consultation interaction data stream refers to the real-time interaction content between customers and consultation agents (human customer service representatives or intelligent customer service robots). This data stream includes at least the following two forms: (1) Real-time consultation text: Text messages sent by the customer one by one in the consultation session window, as well as text messages replied by the agent. The text messages contain unstructured information such as the customer's natural language questions, descriptions of needs, and scenario descriptions.

[0035] (2) Voice-to-text transcription of price inquiries: When customers inquire about prices or make price quotes through the voice channel, the system calls the automatic speech recognition engine to convert the speech stream into text in real time, forming voice-to-text transcription. Since speech recognition may have errors, the system can perform post-processing on the transcribed text, including noise reduction, error correction, and punctuation restoration.

[0036] The real-time consultation interaction data stream can be obtained by subscribing to streaming data through a message queue middleware. The professional knowledge service platform sends consultation messages one by one or in batches to the message queue, and this system, as a consumer, pulls data from the message queue to achieve low-latency data acquisition.

[0037] Step S103: Extract knowledge from the attribute feature data of the customer to obtain a basic customer knowledge graph.

[0038] Knowledge is extracted from the above attribute feature data to construct a basic customer knowledge graph that represents the static profile of customers.

[0039] The aforementioned basic customer knowledge graph includes semantic relationships between customer attributes and customer behaviors, such as knowledge units in the form of triplets like "Customer Zhang San - Age - 35 years old", "Customer Zhang San - Browsing - Car Insurance Product Page", and "Customer Zhang San - Searching - Third-Party Liability Insurance". These knowledge units are integrated to form a structured graph representation.

[0040] Step S104: The real-time consultation interaction data stream is fused with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph.

[0041] The real-time consultation interaction data stream obtained in step S102 is integrated with the basic customer knowledge graph constructed in step S103 through online knowledge fusion, so that the real-time extracted customer consultation intent is semantically associated with the static customer profile, forming a dynamically updated customer intent knowledge graph.

[0042] Step S105: In response to the change in the target customer's intent state in the customer intent knowledge graph meeting the preset trigger conditions, determine the target ad type based on the target customer's current intent tag, and retrieve the corresponding target ad for real-time delivery.

[0043] Step S105 is the decision-making and execution step for ad placement. The system continuously monitors the intent status of each customer in the customer intent knowledge graph. When it detects that a change in the intent status of a customer (i.e., the target customer) meets the pre-set trigger conditions, it automatically executes the ad placement decision.

[0044] Figure 2 This is a flowchart illustrating the basic customer knowledge graph construction process provided in an embodiment of the present invention. Figure 2 As shown, step S103 specifically includes the following sub-steps: Step S1031: Perform data identification on the customer's attribute feature data to distinguish between structured and unstructured data.

[0045] The aforementioned structured data refers to data with a predefined format, such as field values ​​in a database table (login ID, age, gender, region, etc.); unstructured data refers to text data without a predefined format, such as search keywords entered by customers when browsing, and page description text.

[0046] Specifically, data identification refers to determining the type of customer attribute data according to its data storage format and structure: the system examines the storage format and field organization of each attribute data item. If the data is stored in a relational database table with predefined field formats, and each field has a clear data type (such as integer, string, or date) and field meaning (such as "user_id", "age", "city"), and the data values ​​are within a preset range, then it is determined to be structured data. Typical structured data includes: customer login ID, age value, gender enumeration value, region name, device type label, etc.

[0047] Data stored in formats such as free text, log text, or nested JSON strings, without predefined field format constraints, with variable content length, and semantic information implicit in natural language descriptions, is considered unstructured data. Typical unstructured data includes: search keywords entered by customers in search boxes (such as "how much is factory fire insurance?"), page description text viewed by customers (such as page titles "Enterprise Property Insurance Guide"), and text fields in customer interaction logs with the system.

[0048] After data identification is complete, the system routes structured and unstructured data to different processing channels, where the corresponding extraction modules perform subsequent processing. This channel-based processing design fully leverages the advantages of different extraction techniques, improving the accuracy and efficiency of knowledge extraction.

[0049] Step S1032: For structured data, extract customer attribute entities from it using preset mapping rules.

[0050] For structured data, due to its regular format and clear field meanings, the system directly extracts customer attribute entities based on mapping rules.

[0051] The aforementioned predefined mapping rule is a set of predefined correspondences between fields and entity types. This mapping rule defines which field or combination of fields in the database table corresponds to which type of entity in the knowledge graph.

[0052] For example, the database field “user_id” is mapped to a unique identifier attribute of the “customer entity”; The database field "age" is mapped to the "age attribute", the corresponding entity is the "customer entity", and the relation type is "has age"; the database field "gender" is mapped to the "gender attribute", the corresponding entity is the "customer entity", and the relation type is "has gender"; the database field "city" is mapped to the "region entity", the corresponding entity is the "customer entity", and the relation type is "location"; the database field "device_type" is mapped to the "device attribute", the corresponding entity is the "customer entity", and the relation type is "uses device".

[0053] The configuration of the above mapping rules can be manually defined by domain experts, or it can be automatically generated by analyzing database table structure comments and data dictionaries.

[0054] In one possible implementation, the system reads structured data records one by one according to the mapping rules, converts each field value into a corresponding (entity, attribute, attribute value) triple, and completes the extraction of customer attribute entities.

[0055] For example, for a structured data record: {user_id:"U001",age:35,gender:"Male",city:"Chengdu",device_type:"Mobile"}, after processing by mapping rules, the following preliminary knowledge units can be obtained: (Customer entity_U001, with age, 35), (Customer entity_U001, with gender, male), (Customer entity_U001, location, region entity_Chengdu), (Customer entity_U001, using device, Mobile).

[0056] Step S1033: For unstructured data, use named entity recognition technology to extract customer behavior entities, and use relation extraction technology to identify the semantic relationship between customer attribute entities and customer behavior entities.

[0057] For unstructured data, since its content is natural language text and semantic information is implicit in the sentences, it cannot be directly extracted through simple mapping rules and requires in-depth analysis using natural language processing techniques.

[0058] First, the system utilizes named entity recognition (NER) technology to extract customer behavior entities from unstructured text. NER is a sequence labeling task whose goal is to identify entity fragments with specific meanings from a text sequence and label their categories. In this embodiment, the system can employ a deep learning model based on pre-trained language models (such as BERT, RoBERTa, etc.) combined with conditional random field layers as the named entity recognizer. This model is pre-trained on labeled corpora in the insurance field to enable it to recognize entity types related to insurance consultation and customer behavior, including but not limited to: insurance product names (such as "corporate property insurance", "public liability insurance", "employer's liability insurance"), insurance terms (such as "deductible", "insured amount", "insurance period"), risk types (such as "fire risk", "water damage risk", "theft risk"), and behavioral actions (such as "inquiry", "insurance application", "claim", "renewal"), etc.

[0059] The process of named entity recognition is as follows: Unstructured text (such as the search keyword "how much is factory fire insurance") is input into a pre-trained named entity recognition model. The model annotates the text character by character and outputs the start and end positions of entity segments and entity category labels. For example, from "how much is factory fire insurance", "factory fire insurance" is identified as an entity of the "insurance product" category.

[0060] Secondly, based on the extraction of customer behavior entities, the system uses relation extraction technology to identify the semantic relationships between customer attribute entities and customer behavior entities. The goal of relation extraction is to determine whether a predefined semantic relationship exists between two entities in the text and to identify the type of relationship.

[0061] In this embodiment, the system can employ a relation classifier based on a pre-trained language model to extract relations. Unstructured text and its identified entity pairs are taken as input, and the relation classifier outputs the relationship type between these entity pairs. Predefined relation types include, but are not limited to: "Search" (a customer searched for a product), "Browse" (a customer browsed a page type), "Inquire" (a customer asked a question), and "Follow" (a customer followed a topic), etc.

[0062] For example, for the search keyword "how much is factory fire insurance", the system has identified "Customer Entity_U001" (associated with login ID) and "Insurance Product Entity_Factory Fire Insurance", and the relationship extraction model determines the relationship type between these two entities as "Search". For the page description text corresponding to the page browsing behavior, the system identifies "Customer Entity_U001" and "Page Type Entity_Product Introduction Page", and the relationship extraction model determines the relationship type as "Browse".

[0063] It should be noted that for multiple unstructured data entries generated by the same customer (such as multiple searches or browsing multiple pages), the system extracts entities and relationships from each data entry separately, resulting in multiple sets of independent entity-relation pairs. These entity-relation pairs will be integrated in the subsequent knowledge fusion step.

[0064] Step S1034: Organize the extracted customer attribute entities, customer behavior entities, and semantic relationships into preliminary knowledge units in the form of triples.

[0065] Triples are the basic building blocks of knowledge graphs, with a general format of (head entity, relation, tail entity). Head and tail entities are nodes in the knowledge graph, and relations are directed edges connecting the two nodes. In this embodiment, the entities extracted in steps S1032 and S1033 are used as head or tail entities, and attribute relations or semantic association relations are used as relations, all organized into preliminary knowledge units in the form of triples.

[0066] The initial knowledge units are: triples come from different data sources (structured data mapping and unstructured data extraction), and there may be cases where the same entity appears in different triples with different names or identifiers (i.e., entity ambiguity); triples describing the same semantic relationship may be scattered in different batches of data processing results and have not yet been integrated and merged.

[0067] The specific organization method is as follows: For results obtained from structured data mapping, triples are formed with the customer entity as the head entity, attribute relationships as the relations, and attribute values ​​or attribute entities as the tail entities, for example, (Customer_U001, has age, 35) and (Customer_U001, location, Chengdu). For results obtained from unstructured data extraction, triples are formed with the customer entity as the head entity, behavioral semantic relationships as the relations, and behavioral entities or product entities as the tail entities, for example, (Customer_U001, search, factory fire insurance) and (Customer_U001, browse, product introduction page).

[0068] All triples are stored in a temporary knowledge base, awaiting subsequent knowledge fusion processing.

[0069] Step S1035: Based on the customer domain ontology model, perform knowledge fusion on the preliminary knowledge units.

[0070] Since the initial knowledge units originate from different data processing channels, the semantic heterogeneity and entity naming inconsistencies between data sources need to be addressed through knowledge fusion. The purpose of knowledge fusion is to eliminate entity ambiguities from different data sources and integrate scattered semantic relationships to generate semantically consistent and non-redundant integrated knowledge units.

[0071] In this embodiment, the core basis for knowledge fusion is the customer domain ontology model. The customer domain ontology model is a pre-built formal semantic model used to describe customer domain concepts and their relationships. This model is formally expressed using the W3C-recommended ontology description language OWL, and its core components include: (1) Class definition: Defines the conceptual categories involved in the customer domain. For example, define the "Customer" class, "Product" class, "Page" class, "Region" class, "Device" class, "Behavioral Event" class, etc. There can be hierarchical relationships between the classes. For example, "Insurance Product" is a subclass of "Product", and "Auto Insurance Product" and "Non-Auto Insurance Product" are subclasses of "Insurance Product".

[0072] (2) Attribute definition: Defines the association between classes and between a class and a data type value. This includes object attributes (connecting instances of two classes) and data attributes (connecting instances of a class with specific numerical values). For example, define the object attribute "location", with its domain being the "customer" class and its value domain being the "region" class; define the data attribute "has age", with its domain being the "customer" class and its value domain being an integer type.

[0073] (3) Relationship constraints: Define the logical characteristics of attributes, such as transitivity, symmetry, and functionality. For example, the "location" attribute has functionality, meaning that a customer can only have one location.

[0074] (4) Instance attribution: Defines which specific entities belong to which class of instances.

[0075] The knowledge integration process specifically includes the following three stages: (a) Entity Alignment: The goal of entity alignment is to identify multiple entity representations pointing to the same real-world object in the initial knowledge unit and merge them into a unified entity. The system extracts key attributes for each entity (such as login ID, device fingerprint, and behavioral feature vector for customer entities; product name and product code for product entities) and calculates the attribute similarity between entity pairs. Similarity calculation methods can include name similarity based on string edit distance, semantic similarity based on vector space models, or rule-based multi-attribute matching. For entity pairs with similarity exceeding a preset threshold, the system determines them as the same entity, combines their triplets, and uses a unified standard entity identifier.

[0076] For example, if a customer logs in using different devices in different sessions, resulting in two sets of customer entities named with different temporary identifiers, but with the same login ID attribute, the entity alignment module identifies that these two customer entities point to the same customer by matching the login ID and merges them into a unified customer entity named with the login ID.

[0077] (b) Attribute Merging: After entity alignment is completed, the system merges the attributes associated with each aligned entity. For the same attribute (such as age) from different data sources, if the attribute values ​​are consistent, they are retained; if there are conflicting attribute values ​​(such as one data source recording an age of 35, and another recording 36), the final attribute value is determined based on factors such as the data source's credibility weight and the data's freshness timestamp, using a preset conflict resolution strategy. For example, the most recently updated data is given priority, or the data source with a higher credibility weight is given priority.

[0078] (c) Relationship integration: Based on entity alignment and attribute merging, the system aggregates all triple relationships associated with a unified entity, eliminates duplicate relationship triples (for example, records of the same customer searching for the same product multiple times can be retained as multiple independent behavioral events, but the search relationship itself is not repeatedly established), and forms an integrated set of semantic relationships.

[0079] After completing the above three steps, the initial knowledge unit is transformed into an integrated knowledge unit that is semantically consistent, entity-unified, and has clear relationships.

[0080] It should be noted that there is a predefined semantic mapping relationship between the customer domain ontology model and the cross-domain intent ontology model used in the following steps. The core concepts such as "customer," "product," and "behavioral event" defined in the customer domain ontology model are corresponding to the concepts such as "customer attribute," "insurance product," and "consultation behavior" defined in the cross-domain intent ontology model through a pre-defined concept mapping table.

[0081] For example, the "Product" class in the customer domain ontology maps to the "Insurance Product" class in the cross-domain intent ontology, and the "Search" and "Browse" subclasses in the "Behavioral Event" class in the customer domain ontology map to the "Inquiry Behavior" and "Insurance Consultation Behavior" classes in the cross-domain intent ontology, respectively. This cross-ontology concept mapping mechanism provides a bridge for semantic alignment between real-time intent entities and static customer entities in subsequent steps, enabling the fusion of knowledge from two different semantic spaces within a unified semantic framework and avoiding entity linking errors caused by ontology heterogeneity.

[0082] Step S1036: Define the entity types and relationship types of the basic customer knowledge graph based on the customer domain ontology model, fill in the integrated knowledge units as instance data, and construct the basic customer knowledge graph.

[0083] After knowledge integration is completed, the system will carry out the final construction of the basic customer knowledge graph.

[0084] The basic customer knowledge graph uses a graph database for physical storage. Graph databases (such as Neo4j and JanusGraph) organize data in the form of nodes and edges, making them naturally suitable for storing and querying knowledge representations with graph structures. The basic customer knowledge graph is logically divided into two layers: a schema layer and a data layer. (1) Schema Layer: Defined by the customer domain ontology model. The “classes” defined in the ontology model correspond to the “entity types” (i.e., node labels) of the knowledge graph. For example, “customer”, “product”, “page”, and “region” are entity types. The “object attributes” and “data attributes” defined in the ontology model correspond to the “relationship types” (i.e., edge labels) of the knowledge graph. For example, “location”, “search”, “browse”, and “with age” are relationship types. The schema layer specifies which types of nodes and edges are allowed in the graph, which is equivalent to the graph’s “data dictionary” or “structural constraints”.

[0085] (2) Data layer: populated by the integrated knowledge units generated in step S1035. Entity instances in the integrated knowledge units (such as “Customer_U001”, “Factory Fire Insurance”, “Chengdu”) serve as nodes in the data layer and are assigned to the corresponding entity type labels according to their respective ontology classes; triple relations in the integrated knowledge units serve as directed edges in the data layer and are assigned to the corresponding relation type labels according to their relation types.

[0086] The construction process is as follows: The system traverses the integrated set of knowledge units. For each triple (head entity, relation, tail entity), it performs node and edge creation or merging operations in the graph database. If the head entity node or tail entity node already exists (determined by the entity's unique identifier), attributes or relations are appended to it; otherwise, a new node is created and its entity type and attributes are set. For relations, a directed edge of type relation is created, starting from the head entity node and ending at the tail entity node, and the relevant contextual attribute information (such as the timestamp of the action, session ID, etc.) is stored as the edge's attributes.

[0087] For example, for the integrated knowledge unit (customer_U001, search, factory fire insurance), the system creates an edge of type "search" between the customer_U001 node and the factory fire insurance node, and records attribute information such as the timestamp of the search and the source channel of the search on the edge.

[0088] Once constructed, the basic customer knowledge graph becomes the structured semantic representation of the system's static customer profile. In this graph, various customer attributes (age, gender, region, device, etc.) are recorded as node attributes, customer behaviors (search, browsing, clicking, etc.) are recorded as relationships between nodes, and interactions between customers and objects such as products and pages are recorded as relationship edges. This structured semantic representation supports complex semantic queries using flexible knowledge graph query languages ​​(such as SPARQL and Cypher), such as "query all customers who have searched for factory fire insurance and whose location is Chengdu".

[0089] Figure 3 This is a flowchart illustrating the customer intent knowledge graph construction process provided in an embodiment of the present invention. Figure 3 As shown, step S104 specifically includes the following sub-steps: Step S1041: Perform streaming knowledge extraction on the real-time consultation interaction data stream.

[0090] Since the real-time consultation and interaction data stream is continuously arriving streaming data rather than batch static data, the system needs to use a streaming processing mechanism to extract knowledge from it in real time.

[0091] In this embodiment, the system deploys a streaming processing engine (such as Apache Flink, Kafka Streams, etc.) to process the real-time consultation text and / or quotation speech-transcribed text obtained from the message queue, either line-by-line or in micro-batch. The processing procedure is as follows: First, the system preprocesses each newly arrived text message, including text cleaning (removing irrelevant characters, emoticons, HTML tags, etc.), word segmentation, and part-of-speech tagging.

[0092] Secondly, the system calls pre-trained intent entity extraction and intent classification models to analyze the pre-processed text. The intent entity extraction model is used to identify entities related to the customer's consultation intent from the text, including but not limited to: the names of insurance products mentioned by the customer (such as "all risks of property insurance" and "machinery breakdown insurance"), the types of risks the customer is concerned about (such as "fire", "water damage", and "theft"), the insurance elements the customer inquires about (such as "premium", "coverage amount", and "deductible"), and the business scenarios described by the customer (such as "new factory insurance" and "renewal of old policy"). The intent entity extraction model can also be implemented using a sequence labeling model based on a pre-trained language model, and its training corpus comes from historical consultation dialogue annotation data accumulated by the professional knowledge service platform.

[0093] The intent classification model is used to identify the category of the inquiry intent expressed in a text message, i.e., the inquiry focus. The category labels for inquiry focuses include, but are not limited to: "price inquiry," "coverage inquiry," "claims process inquiry," "underwriting conditions inquiry," "product comparison inquiry," and "renewal inquiry." The intent classification model can be implemented using a text classification model based on a pre-trained language model, outputting one or more intent category labels and corresponding confidence scores for each inquiry message.

[0094] Simultaneously, the system also needs to extract the contextual semantic relationship between the intent entity and the consultation focus. Contextual semantic relationship refers to how the intent entity and the consultation focus are related within the context of the current consultation session. For example, if a customer says, "I want to ask how much factory fire insurance costs," then "factory fire insurance" is the intent entity, and "price inquiry" is the consultation focus; the contextual semantic relationship between them is "inquiry—price." By analyzing syntactic dependency relations and semantic role labeling results, the system determines the type of semantic association between the intent entity and the consultation focus, forming a triple structure of (intent entity, semantic relationship, consultation focus).

[0095] The intent entities, consultation focus, and contextual semantic relationships extracted in real time from a single or multiple adjacent consultation messages collectively constitute a real-time intent knowledge unit. A real-time intent knowledge unit is an immediate knowledge representation of the intent revealed in the customer's current consultation behavior. As the consultation session continues, real-time intent knowledge units are continuously generated, forming a sequence of real-time intent knowledge units.

[0096] Step S1042: Based on the cross-domain intent ontology model, integrate the real-time intent knowledge units with the basic customer knowledge graph.

[0097] Since the basic customer knowledge graph is built on a customer domain ontology model, its entity types, relationship types, and semantic constraints are all defined within this ontology framework; while the real-time intent knowledge unit is extracted from consultation interaction text, and its semantic expression follows the intent semantic system of the consultation domain. There is a semantic space difference between the two. To address this issue, step S1042 introduces a cross-domain intent ontology model as a semantic bridge. This cross-domain intent ontology model not only defines the semantic hierarchy and relationship constraints between preset knowledge categories, but also establishes a connection with the customer domain ontology model through the aforementioned concept mapping table, thereby supporting cross-semantic space entity links and semantic relationship establishment between intent entities in the real-time intent knowledge unit and customer entities in the basic customer knowledge graph. This dual-layer ontology collaborative knowledge fusion method overcomes the limitation that a single ontology model cannot simultaneously cover static customer profiles and dynamic consultation intents, achieving cross-domain semantic alignment.

[0098] Cross-domain intent ontology models are another type of ontology model, distinct from customer domain ontology models. Customer domain ontology models focus on defining the conceptual system of static customer attributes and behaviors, while cross-domain intent ontology models focus on defining the semantic hierarchy and relational constraints between cross-domain knowledge categories such as customer intent, insurance products, risk types, and consulting behaviors.

[0099] Specifically, the cross-domain intent ontology model is formally expressed using the OWL ontology description language, defining the following predefined knowledge categories and their semantic relationships: (1) Customer Attributes: Categories that describe the basic characteristics of customers, including concepts such as age, gender, region, occupation, and industry.

[0100] (2) Insurance Products Category: This describes the categories of insurance product classification system, including two major subcategories: motor insurance and non-motor insurance. Non-motor insurance is further subdivided into subcategories such as property insurance, liability insurance, credit insurance, and guarantee insurance. Each subcategory is further subdivided into specific types of insurance (such as all-risk property insurance, machinery breakdown insurance, and public liability insurance).

[0101] (3) Risk Intent Category: This category describes the risk concerns revealed by clients during consultations, including property risk protection intent, personal risk protection intent, liability risk avoidance intent, and business interruption risk protection intent. There is a hierarchical relationship between risk intent categories; for example, "fire risk protection intent" is a subcategory of "property risk protection intent".

[0102] (4) Consultation behavior category: describes the types of customer behavior during the consultation process, including inquiry behavior, price comparison behavior, insurance consultation behavior, claims consultation behavior, renewal behavior, etc.

[0103] The cross-domain intent ontology model also defines the semantic hierarchy and relational constraints between the aforementioned knowledge categories. For example, relational constraints are defined as follows: "customer" and "risk intent" are associated through the "having intent" relationship; "customer" and "insurance product" are associated through the "consulting product" relationship; "risk intent" and "insurance product" are associated through the "insurable" relationship; and "consulting behavior" and "insurance product" are associated through the "product-specific" relationship.

[0104] Figure 4 This is a flowchart illustrating the knowledge fusion process provided in an embodiment of the present invention. Figure 4 As shown, step S1042 specifically includes the following sub-steps: Step S1042a: Link the intent entities in the real-time intent knowledge unit with the customer entities in the basic customer knowledge graph to identify the customer entities that point to the same real-world entity as the intent entities.

[0105] The goal of entity linking is to identify which customer entity in the real-time intent knowledge unit points to the same real-world subject as the intent entity in the basic customer knowledge graph, thereby establishing a correspondence between real-time intent information and existing customer profiles.

[0106] The customer entity is a node representing the customer in the basic customer knowledge graph. This node is created when constructing the basic customer knowledge graph in step S103 and records the customer's historical attribute information and behavioral information.

[0107] The intent entity is a node in the real-time consultation interaction data stream that represents the customer's consultation intent. This node is extracted from the real-time consultation message in step S1041 and represents the customer's identity and intent in this consultation session.

[0108] The real-world entity is the customer who generates the real-time consultation interaction data stream; that is, the natural person in the physical world operating the terminal device and interacting with the consultation system. Both the customer entity and the intent entity are digital representations of this real-world entity at different data layers within the system.

[0109] The specific execution process of entity linking is as follows: The first step is to extract the attribute features of the intent entity. Although the intent entity is extracted from a single consultation, it carries attribute information related to the customer's identity. This attribute information can be obtained from the context metadata of the consultation session, including but not limited to: the session identifier that initiated the consultation, the customer's login status token, device fingerprint, IP address, and consultation channel source (such as which page entry it came from).

[0110] The second step is to extract the attribute features of candidate customer entities from the basic customer knowledge graph. The system uses key identifiers (such as login IDs) from the attribute features of the intent entity as query criteria to retrieve a set of candidate customer entities from the graph. If the intent entity carries a clear login ID, the candidate set is usually narrowed down to a unique customer entity; if the identity information in the intent entity is incomplete (e.g., a customer initiating a consultation as a visitor), a broader attribute matching process is needed to determine the candidate set.

[0111] The third step is to calculate the attribute similarity between the intent entity and each candidate customer entity. Similarity calculation can combine multiple dimensions of features for weighted scoring, such as: exact match score for identity identifiers, similarity score for device fingerprints, geographic match score for IP address location, and consistency score for behavioral time sequence (the shorter the time interval between the inquiry and the most recent browsing behavior, the higher the score). The scores of each dimension are weighted and summed according to preset weights to obtain the comprehensive similarity score.

[0112] The specific execution process of entity linking is as follows: Extract the attribute features of the intent entity.

[0113] Although the intent entity is extracted from a single consultation, it carries attribute information related to the customer's identity. This attribute information can be obtained from the context metadata of the consultation session, including but not limited to: the session ID that initiated the consultation, the customer's login state token (Auth Token, from which the login ID can be parsed), the device fingerprint (a unique identifier generated by a combination of device type, operating system version, screen resolution, browser type and version, etc.), the IP address (which can be parsed to determine the location and carrier), and the source of the consultation channel (such as which page entry it came from, which search keyword it was introduced by), etc.

[0114] Extract the attribute features of candidate customer entities from the basic customer knowledge graph.

[0115] The system uses key identifiers (such as login ID) in the attribute characteristics of intent entities as query conditions to retrieve a set of candidate customer entities in the graph. If the intent entity carries a clear login ID, the candidate set is usually narrowed down to a unique customer entity; if the identity information of the intent entity is incomplete (such as a customer initiating an inquiry as a visitor without carrying a login token), a broader attribute matching is needed to determine the candidate set, typically selecting all recently active customer entities in the graph (such as those with behavioral records in the last 24 hours) as candidates.

[0116] Calculate the attribute similarity between the intent entity and each candidate customer entity.

[0117] Similarity calculation employs a weighted scoring system that integrates features from multiple dimensions. The weights for each dimension are pre-set based on the distinguishability and reliability of each attribute in identity recognition. The weights and scoring methods for each dimension are as follows: Identity identification dimension (highest weighting, e.g., preset weight 40%): Includes unique identification information such as login ID, customer account, registered mobile phone number or email address. If the login ID of the intent entity is exactly the same as the login ID of the candidate customer entity, this dimension scores full marks (e.g., 1.0); if they are inconsistent or either party is missing this information, the score is 0.

[0118] Device fingerprint dimension (with secondary weight, such as a preset weight of 25%): Device type, operating system version, browser type, screen resolution, and other information are concatenated into a device feature string. The score for this dimension is calculated by the edit distance similarity or Jaccard similarity between two device feature strings. Complete device similarity scores 1.0; partial similarity is scored proportionally.

[0119] IP address location dimension (weighted again, e.g., preset weight of 20%): Extract the location information corresponding to the IP address (accurate to the city level). If two IP addresses have the same location, the score is 1.0; if they have different locations but belong to the same province, the score is 0.5; otherwise, the score is 0.

[0120] Behavioral Sequence Coherence Dimension (weighted again, e.g., preset weight of 15%): Calculate the time interval between the consultation initiation time corresponding to the intent entity and the time of the candidate customer entity's most recent behavior record (e.g., the time of the most recent browsing, search, or consultation). The shorter the time interval, the higher the probability that the two behavioral events belong to the same customer's coherent behavior. If the time interval is within 1 minute, the score is 1.0; between 1 minute and 5 minutes, the score is 0.8; between 5 minutes and 30 minutes, the score is 0.5; between 30 minutes and 2 hours, the score is 0.3; and more than 2 hours, the score is 0.1.

[0121] The scores of each dimension are weighted and summed according to their corresponding preset weights to obtain the comprehensive similarity score. The comprehensive similarity score ranges from 0 to 1.0, with a value closer to 1.0 indicating a higher probability that the two entities refer to the same real-world subject.

[0122] Specific calculation example: Suppose customer "Zhang San" logs into the website and browses the car insurance product page using an Android phone (device type = Mobile, operating system = Android 13, browser = Chrome 120) under a network environment with the IP address "Chengdu Telecom", and then initiates an inquiry. His login ID is "U001". The intent entity extracted by the system from the real-time inquiry interaction data stream carries the following attributes: Login ID = U001, Device fingerprint = "Mobile_Android13_Chrome120_1080x2340", IP location = "Chengdu", Inquiry initiation time = "2024-01-15 14:32:18".

[0123] In the basic customer knowledge graph, there is a candidate customer entity “Customer_U001” with the following attributes: Login ID=U001, Device fingerprint=“Mobile_Android13_Chrome120_1080x2340” (which is completely consistent with the device fingerprint of the intent entity), IP location=“Chengdu”, and Recent behavior time=“2024-01-15 14:31:45” (33 seconds after the inquiry was initiated).

[0124] The scores for each dimension are calculated as follows: Identity identification dimension: Both login IDs are U001, completely identical, score 1.0, weight 40%, weighted score 0.40; Device fingerprint dimension: Device feature strings are completely identical, score 1.0, weight 25%, weighted score 0.25; IP address location dimension: Both are "Chengdu", score 1.0, weight 20%, weighted score 0.20; Behavioral temporal continuity dimension: Time interval 33 seconds (within 1 minute), score 1.0, weight 15%, weighted score 0.15. Overall similarity score = 0.40 + 0.25 + 0.20 + 0.15 = 1.0.

[0125] The fourth step involves identifying the candidate customer entity with the highest overall similarity score exceeding a preset threshold (e.g., 0.7) as the customer entity pointing to the same real-world entity as the intent entity. In the example above, the only candidate customer entity has an overall similarity score of 1.0, exceeding the threshold of 0.7. Therefore, the system determines that the intent entity and "customer_U001" point to the same real-world entity, "Zhang San". If the similarity scores of all candidate customer entities are below the threshold, it may indicate that the customer is a new customer (not yet documented in the customer resource database). The system can then choose to create a new customer entity node and gradually enrich its attribute information in subsequent actions.

[0126] The fifth step involves identifying the candidate customer entity with the highest overall similarity exceeding a preset threshold as the customer entity pointing to the same real-world entity as the intent entity. If the similarity of all candidate customer entities is below the threshold, it may indicate that the customer is a new customer (not yet documented in the customer database), and the system may choose to create a new customer entity node.

[0127] For example, after logging into the website, customer "Zhang San" initiates a consultation. The login state token of the consultation session contains his login ID "U001". The intent entity extracted by the system from the real-time consultation interaction data stream carries the attribute "Login ID=U001". In the basic customer knowledge graph, there already exists a customer entity node whose attribute "Login ID" is also "U001". By precisely matching the "Login ID" attribute, the system calculates that the similarity between the two is 100%, determining that the intent entity and the customer entity node "Customer_U001" point to the same real-world entity "Zhang San".

[0128] Step S1042b: Based on the relational constraints defined by the cross-domain intent ontology model, establish real-time semantic associations between intent entities and identified customer entities. The type of real-time semantic association is determined by the relational constraints.

[0129] After the entity link determines the correspondence between the intent entity and the customer entity, the system needs to establish a real-time semantic association between the two so that the graph can express the dynamic semantic meaning of "the customer currently has a certain consultation intent".

[0130] The establishment of real-time semantic associations is based on the relational constraints defined by the cross-domain intent ontology model. Specifically, the system selects the most matching relational type from the relational types defined by the cross-domain intent ontology model as the type of real-time semantic association based on the semantic content extracted from the real-time intent knowledge unit.

[0131] The relation constraints predefine various relation types and their applicable conditions. For example, if a real-time intent knowledge unit contains a customer's inquiry about an insurance product, the relation type is selected as "Inquire about Product"; if a real-time intent knowledge unit contains a customer's inquiry about price, the relation type is selected as "Inquire about Focus"; if a real-time intent knowledge unit contains a customer's description of a risk scenario, the relation type is selected as "Focus on Risk".

[0132] The system performs the following operations in the graph database: starting from the identified customer entity node and ending at the intent entity node (if the intent entity does not yet exist as an independent node in the graph, the intent entity node is created first and its type label is set), a directed edge is created, with the edge's type label being a defined relation type. This directed edge is the physical representation of the real-time semantic association in the graph.

[0133] Step S1042c: The consultation focus and contextual semantic relationship associated with the intent entity are added to the customer intent knowledge graph as attribute information of real-time semantic association, resulting in a dynamically updated customer intent knowledge graph.

[0134] After establishing edges with real-time semantic associations, the system also needs to store the detailed semantic information extracted in step S1041 as attributes of the edge to enrich the expressive power of the graph.

[0135] Specifically, the system stores the consultation focus (such as intent category labels like "price consultation" and "coverage consultation") as the value of the "consultation focus" attribute field of the real-time semantic association edge. The contextual semantic relationship (such as "inquiry-price" and "concern-coverage") is stored as the value of the "contextual semantic" attribute field of the real-time semantic association edge. Furthermore, metadata such as the timestamp of the consultation and the original text of the consultation message can be stored as additional attributes of the edge.

[0136] For example, in the scenario where customer "Zhang San" inquires about "how much is factory fire insurance", the system establishes a real-time semantic association edge of type "consultation product" between the "customer_U001" node and the "intent entity_001" node, and records the attributes on this edge: consultation focus = "price consultation", context semantics = "inquiry - price", product = "factory fire insurance", timestamp = "2024-01-15 14:32:18".

[0137] After steps S1042a to S1042c, the system completes the knowledge fusion processing for a single real-time consultation event. As the consultation session continues, new real-time consultation messages are constantly generated. The system repeats the above process for each message or group of messages, continuously fusing newly extracted intent entities, consultation focus, and contextual semantic relationships into the customer intent knowledge graph. Therefore, the customer intent knowledge graph is a dynamically updated graph structure, whose nodes and edges continuously grow and evolve as the customer's consultation behavior progresses.

[0138] After each knowledge fusion operation, the system obtains the latest customer intent knowledge graph up to the current moment. Compared with the basic customer knowledge graph, this graph adds intent nodes, semantic association edges, and attribute information related to the customer's real-time consultation behavior, realizing a dynamic characterization of the customer's intent state.

[0139] Figure 5 This is a flowchart illustrating the advertising placement decision-making process provided in an embodiment of the present invention. Figure 5 As shown, step S105 specifically includes the following sub-steps: Step S1051: Monitor the status changes of the intent nodes corresponding to the target customer in the customer intent knowledge graph.

[0140] The system continuously monitors the status of each customer whose intent-related nodes and edges have been established in the customer intent knowledge graph.

[0141] The object of monitoring is the state change of the intent node corresponding to the target customer.

[0142] The status information of an intent node includes, but is not limited to: the consultation focus tag associated with the intent node, the risk intent level corresponding to the intent node, the number of real-time semantic association edges related to the intent node (reflecting consultation activity), and the last update time of the intent node.

[0143] The state changes of the intent node are triggered by the knowledge processing operations performed by the system on the customer intent knowledge graph.

[0144] The aforementioned knowledge processing operations are the process by which the system performs in-depth analysis and intelligent reasoning on the graph. The purpose is to deduce implicit, higher-level intention cognition from the existing explicit knowledge in the graph. The knowledge processing operations can be triggered after each knowledge fusion and graph update, or they can be executed in batches according to a preset time period (e.g., every 30 seconds).

[0145] Knowledge processing operations include at least one of the following two mechanisms: (1) Intention reasoning based on ontology rules.

[0146] Intent reasoning based on ontology rules is a deductive reasoning method that uses predefined logical rules to deduce new facts or update the state of existing facts from existing facts in a graph.

[0147] The specific process is as follows: The system acquires the intent entities, consultation focuses, and real-time semantic association information associated with the target customer from the customer intent knowledge graph, and uses this information as the input fact base for the inference engine. The inference engine loads a pre-defined ontology rule base, and the rules in the rule base are written using the Semantic Web rule language SWRL. Each rule has a "condition-conclusion" structure, indicating that when the input facts satisfy the condition part, the conclusion part of the facts is derived.

[0148] For example, the following inference rules are presupposed: Rule 1: If a customer sends more than a preset threshold number of inquiries about the same insurance product category within a preset time window, and all inquiries focus on "price inquiries", then the customer's risk intent level will be updated to "price comparison intent - strong".

[0149] Rule 2: If a customer's inquiry message mentions both the keywords "factory" and "fire", and the customer's location is a high-risk area for fires according to historical claims data, then the customer's risk intent level will be upgraded from the default level to "Property Risk Protection Intent - High Concern".

[0150] Rule 3: If a customer's inquiry focus is "renewal inquiry", and the basic customer knowledge graph records that the customer has previously purchased an insurance product and the insurance period of that product is about to expire, then the customer's intent tag will be updated to "renewal intent".

[0151] The inference engine matches each rule one by one. When the condition part of a rule is met, the execution of the conclusion part is triggered, updating the status attributes (such as risk intent level, intent tag, etc.) of the corresponding intent node in the customer intent knowledge graph.

[0152] (2) Intent relationship completion based on graph embedding.

[0153] Intent-relationship completion based on graph embedding is an inductive reasoning method that uses knowledge graph embedding technology to map entities and relationships in the graph to a low-dimensional continuous vector space. Through vector operations, it discovers potential, unrecorded relationships in the graph.

[0154] The training data is sourced from the current state of the customer intent knowledge graph. Specifically, the system exports all existing triples (head entity, relation, tail entity) in the customer intent knowledge graph from the graph database as positive training samples. These triples include: customer attribute triples (e.g., customer_U001—with age—35), customer behavior triples (e.g., customer_U001—search—factory fire insurance), and real-time semantic association triples added after knowledge fusion in step S104 (e.g., customer_U001—consulting product—intent entity_001, intent entity_001—consulting focus—price consultation). Simultaneously, the system employs a negative sampling strategy to construct negative training samples. Specifically, for each positive triple, its head or tail entity is randomly replaced, generating relation triples that do not currently exist in the graph as negative examples.

[0155] Model Training and Update Mechanism: After the customer intent knowledge graph completes a knowledge fusion update in step S104, or after new triples are generated by intent reasoning based on ontology rules in step S1051, the graph structure changes. If the system detects that the number of newly added or modified triples in the graph exceeds a preset incremental update threshold (e.g., more than 50 new triples), it triggers incremental training of the graph embedding model. Incremental training uses the model parameters obtained from the previous training round as initial values, fine-tuning the parameters only on the newly added triples and a randomly selected portion of historical triples, without retraining all data from scratch, thus ensuring the real-time nature of model updates. If the number of newly added triples does not exceed the threshold, the current model continues to be used for prediction.

[0156] The specific execution of intent relationship completion: The system predicts potential but not yet established intent relationships in the graph. The prediction method is as follows: For any customer entity and any intent entity or product entity in the graph, their vectors are substituted into the scoring function of the embedded model to calculate a score indicating a specific relationship (e.g., "has intent," "consults on products"). Relationships with scores exceeding a preset completion confidence threshold (e.g., scores higher than 0.8) are considered potential relationships. The system completes missing intent entities or real-time semantic relationships in the graph based on the prediction results. For example, if the model scores the "consults on products" relationship between customer_U001 and the business interruption insurance product entity at 0.85, exceeding the threshold, but the graph has not yet recorded any association between this customer and "business interruption insurance," the system adds a relationship edge of type "potential intent" to the graph and updates the customer's intent state accordingly (e.g., updating the intent state to "potential protection planning intent"). Completion operations may reveal potential intentions that customers have not yet explicitly expressed but are implied in behavioral patterns and similar customer group characteristics, providing a more forward-looking basis for advertising decisions.

[0157] The specific process is as follows: First, the system uses knowledge graph embedding models (such as Trans E, Trans R, Rotat E, etc.) to train the customer intent knowledge graph, learning each entity and each relationship in the graph as a low-dimensional vector representation. After the embedding model is trained, the potential relationships between entities can be predicted through vector operations.

[0158] Secondly, the system predicts potential but not yet established intent relationships within the graph. The prediction method is as follows: for any customer entity and any intent entity or product entity in the graph, an embedding model is used to calculate a score indicating a specific relationship (such as "having intent" or "consulting about products") between the two. Relationships with scores exceeding a preset threshold are considered potential relationships.

[0159] Then, the system completes the missing intent entities or real-time semantic associations in the graph based on the prediction results. For example, if the model predicts that a customer may have a potential intent (high score) for "business interruption insurance", but the graph has not yet recorded any association between the customer and "business interruption insurance", the system adds a "potential intent" type relationship edge to the graph and updates the customer's intent status accordingly (such as updating the intent status to "potential protection planning intent").

[0160] Finally, the system updates the target customer's intent state based on the completed graph. The completion operation may reveal potential intents that customers have not yet explicitly expressed but are implied in behavioral patterns and similar customer group characteristics, providing a more forward-looking basis for advertising decisions.

[0161] Through the above knowledge processing operations, the status of the target customer's corresponding intent node in the customer intent knowledge graph is dynamically updated.

[0162] The two knowledge processing mechanisms mentioned above are complementary: ontology-based intention reasoning can utilize the prior knowledge of domain experts for precise logical deduction, but it is limited by the coverage of existing rules and struggles to discover new intention patterns not defined by the rules; graph embedding-based intention relation completion can learn implicit statistical associations from large-scale graph structures and discover potential intention relationships, but it lacks interpretability and may introduce noise. This invention combines the two: on the one hand, the conclusions of rule-based reasoning can serve as positive examples to supplement the training of the graph embedding model, improving the semantic expressive power of the embedding model; on the other hand, the potential intention relationships discovered by graph embedding completion, after confidence screening, can be submitted as candidate rules to domain experts for review, achieving the inductive discovery of new rules. This mutually reinforcing mechanism enables the customer intention knowledge graph to balance accuracy and comprehensiveness, achieving knowledge evolution during continuous operation.

[0163] Step S1052: When the change in the target customer's intent state is detected to meet the preset triggering conditions, extract the target customer's current intent tag.

[0164] The system continuously compares changes in the target customer's intent with preset trigger conditions. Preset trigger conditions are decision thresholds for determining whether to trigger ad placement. The principle behind setting these conditions is to place ads when the customer's intent is clear enough and their purchase intention is strong enough, avoiding interrupting too early or missing the opportunity too late.

[0165] The preset trigger conditions include at least one of the following: (a) The real-time inquiry frequency of the target customer exceeds the preset threshold. The inquiry frequency reflects the customer's activity level and the urgency of their needs. If the number of messages sent by the customer to the inquiry system within the preset monitoring time window exceeds the preset frequency threshold, it indicates that the customer currently has a strong need for information acquisition, which is an appropriate time to place an advertisement.

[0166] (b) The customer intent knowledge graph indicates that the target customer's risk intent level has upgraded from Level 1 to Level 2, where Level 1 is lower than Level 2. Risk intent level is a quantified or graded representation of a customer's degree of risk concern, for example, it can be divided into three levels: "moderate concern," "moderate concern," and "high concern." When knowledge processing triggers an upgrade in risk intent level from a low level to a high level, it indicates a significant increase in the customer's risk awareness or purchase intention. At this time, advertising related insurance products has a high acceptance rate.

[0167] (c) The target customer made a clear inquiry about the price within the preset time window. Inquiry about the price is the most direct signal of a customer's purchase intention. When the system detects from the real-time consultation interaction data stream that a customer has submitted a formal inquiry request (such as filling out an inquiry form, explicitly asking "how much does this insurance cost?"), it means that the customer has moved from the information gathering stage to the price decision stage. At this time, placing corresponding product or promotional advertisements has the highest conversion rate.

[0168] When at least one of the above triggering conditions is detected to be met, the system extracts the current intent tag corresponding to the target customer.

[0169] The current intent tag is a tag that the system marks for customers during the knowledge processing process, which represents their current main intent type, including but not limited to protection planning intent (the customer is seeking a comprehensive risk protection solution), price comparison intent (the customer is comparing the prices of multiple products), renewal intent (the customer has an insurance policy that is about to expire and needs to be renewed), risk control consultation intent (the customer is concerned about risk management and prevention measures), etc.

[0170] The aforementioned preset threshold is the threshold for the number of consultation messages within a preset time window, such as more than 3 consultation messages within 5 minutes.

[0171] The aforementioned customer intent knowledge graph indicates that the risk intent level of the target customer has been upgraded from level one to level two; the first level is lower than the second level; wherein, the first level is the "general concern" level and the second level is the "high concern" level; the preset threshold is the judgment threshold for the risk intent level upgrade, that is, when the risk intent level attribute value of the target customer's intent node changes from the first level to the second level after the knowledge processing operation, it is determined that the triggering condition is met.

[0172] The aforementioned target customers generated inquiry and quotation behavior within a preset time window; wherein, the preset threshold is a time length threshold of the preset time window, such as within 10 minutes; the inquiry and quotation behavior includes submitting an inquiry form, sending an inquiry message containing the intention to inquire, or triggering the inquiry and quotation function entry.

[0173] Step S1053: Determine the target ad type corresponding to the current intent tag based on the preset mapping relationship between intent tags and ad types.

[0174] The system maintains a pre-defined mapping table that defines the correspondence between various intent tags and ad types. The mapping is set based on business experience and data analysis, guiding customers at different intent stages to the most relevant ad content.

[0175] An example mapping relationship is as follows: Protection planning intent - product combination recommendation advertisement (recommending multi-insurance combination plans); Price comparison intent - promotional advertisement (highlighting price advantages and discount information); Renewal intent - renewal reminder advertisement (reminding policy expiration and renewal process); Risk control consultation intent - risk management service advertisement (recommending value-added services such as risk assessment and disaster prevention and loss mitigation). The system uses the current intent tag extracted in step S1052 as the query key to find the corresponding target advertisement type in the mapping relationship table.

[0176] Step S1054: Retrieve target ads that match the target ad type from the ad resource library and deliver them to the target customers in real time.

[0177] The ad resource library is a pre-built ad creative management system that stores ad creatives and their metadata for different ad types. Each ad creative must contain at least the following information: ad type tag, ad title, ad copy, Uniform Resource Locator (URL) for ad image or video creatives, target redirect link, and applicable delivery channels (web banner, pop-up, sidebar, push notification, etc.).

[0178] The system uses the target ad type determined in step S1053 as the search criteria to filter matching ad creatives from the ad resource library. If there are multiple matching results, one or more target ads can be selected according to the priority weight of the ad creatives, historical click-through rate, or a preset carousel strategy.

[0179] Real-time ad delivery methods can include: embedding ad slots and loading target ad content within the customer's currently viewed website page; displaying target ads as pop-ups on the page; and sending ad summary information via browser push notifications. During delivery, the system records delivery logs, including delivery time, customer ID, ad ID, and delivery channel, for subsequent performance evaluation and data feedback.

[0180] Furthermore, in one embodiment, after real-time ad delivery, the method further includes a data closed-loop optimization step: the system acquires customer conversion behavior data returned by the customer growth management subsystem, including click behavior after ad exposure, landing page browsing behavior after clicking, form submission or order generation behavior after browsing, and final payment completion behavior, etc., conversion link data. The system treats this customer conversion behavior data as new customer behavior data, writes it into the customer resource library through the data feedback channel, and triggers incremental updates to the basic customer knowledge graph and customer intent knowledge graph. During incremental updates, the system updates the attributes and relationships of the corresponding nodes in the graph through the processing flow of steps S103 and S104, so that the knowledge graph continuously reflects the latest status of the customer. This closed-loop mechanism ensures that the effect of each ad delivery can be fed back to the knowledge graph, continuously optimizing the depth of customer understanding and the accuracy of ad delivery strategies.

[0181] It is important to emphasize that the real-time decision-making chain of this invention consists of the acquisition of real-time consultation data stream (step S102), streaming knowledge extraction (step S1041), online knowledge fusion (step S1042), knowledge processing and intent state update (step S1051), and advertising placement decision (steps S1052-S1054). There is a strict temporal coordination relationship between each link in this chain: after the real-time consultation data stream arrives, the system needs to complete knowledge extraction, fusion, processing, and intent state determination within a preset low-latency window to meet the response time requirements of real-time advertising placement. Compared to offline batch processing, this real-time chain design allows the system to complete advertising push before the customer's consultation intent has subsided, significantly improving the timeliness of advertising and customer experience. Furthermore, customer conversion behavior data after placement is fed back to the customer resource database through a feedback channel, triggering incremental updates to the knowledge graph, forming a closed-loop system of "real-time perception—instant decision-making—effect feedback—continuous evolution." Each link forms an organic whole driven by data flow, rather than operating independently. Example 2

[0182] Figure 6 This is a schematic diagram illustrating the structure of a targeted advertising system based on a customer resource database, as provided in an embodiment of the present invention. Figure 6 As shown, the system includes: The attribute data acquisition module 601 is used to extract customer attribute feature data from the customer resource database in response to the login behavior triggered by the customer accessing the website; the customer attribute feature data includes the customer's login information and browsing information.

[0183] The real-time data acquisition module 602 is used to acquire real-time consultation interaction data stream in response to real-time consultation events. The real-time consultation interaction data stream includes at least real-time consultation text and / or transcribed text of consultation quotes.

[0184] The knowledge extraction module 603 is used to extract knowledge from customer attribute feature data to obtain a basic customer knowledge graph; wherein, the basic customer knowledge graph includes the semantic relationship between customer attributes and customer behavior.

[0185] The knowledge fusion module 604 is used to fuse real-time consultation interaction data streams with basic customer knowledge graphs to obtain dynamically updated customer intent knowledge graphs.

[0186] The advertising delivery module 605 is used to respond to changes in the target customer's intent state in the customer intent knowledge graph that meet preset trigger conditions, determine the target ad type based on the target customer's current intent tag, and retrieve the corresponding target ad for real-time delivery.

[0187] The specific functions and implementation methods of the above modules have been described in detail in Embodiment 1 and will not be repeated here. In actual product implementation, the above modules can be software functional modules deployed in a server, executed by the processor calling computer programs in memory to achieve the corresponding functions. The processor of this system can be a chip with computing capabilities such as a central processing unit, a graphics processing unit, or a tensor processor, and the memory can be a storage medium such as random access memory, read-only memory, or flash memory. Example 3

[0188] Figure 7 This is a schematic diagram illustrating the structure of an advertising precision targeting platform based on a customer resource database, provided as an embodiment of the present invention. Figure 7 As shown, the platform includes: The professional knowledge service subsystem 701 is used to provide professional knowledge training and quotation services, generate real-time consultation events, and output real-time consultation interaction data streams.

[0189] The Customer Growth Management Subsystem 702 is used to provide customer growth management services covering customer acquisition, operation, conversion, and management, and to transmit customer conversion behavior data back.

[0190] And the targeted advertising system 703 as described in Example 2.

[0191] The targeted advertising system 703 establishes data connections with the professional knowledge service subsystem 701 and the customer growth management subsystem 702 to receive real-time consultation interaction data streams and customer conversion behavior data. These data connections can be established via internal networks, virtual private clouds, or secure hypertext transfer protocol interfaces to ensure the security and real-time nature of data transmission.

[0192] In one embodiment, the professional knowledge service subsystem 701 is deployed in the first operations center, such as the Chengdu operations center. This center, based on non-motor insurance professional knowledge, builds a professional knowledge training platform and a consulting platform, providing non-motor insurance professional knowledge sharing and quotation services nationwide. The customer growth management subsystem 702 is deployed in the second operations center, such as the Beijing operations center. This center helps brands and merchants build a customer growth management system integrating customer acquisition, operation, conversion, and management based on a full-domain digital growth platform. The advertising precision targeting system 703 can be flexibly deployed in the server clusters of the first and second operations centers, or independently deployed on a cloud service platform, interacting with the above two subsystems via the network.

[0193] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by computer program instructions and related hardware. The computer program can be stored in a computer-readable storage medium, including read-only memory, random access memory, disk, optical disk, etc.

[0194] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0195] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for precise advertising delivery based on a customer resource database, characterized in that, The method includes: In response to a login action triggered by a customer visiting the website, the system extracts the customer's attribute feature data from the customer resource database; the customer attribute feature data includes the customer's login information and browsing information. In response to a real-time consultation event, a real-time consultation interaction data stream is acquired, wherein the real-time consultation interaction data stream includes at least real-time consultation text and / or transcribed text of consultation quotes; Knowledge extraction is performed on the attribute feature data of the customers to obtain a basic customer knowledge graph; wherein, the basic customer knowledge graph includes the semantic association between customer attributes and customer behavior; The real-time consultation interaction data stream is fused with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph. In response to a change in the target customer's intent state in the customer intent knowledge graph that meets a preset trigger condition, the target ad type is determined based on the target customer's current intent tag, and the corresponding target ad is retrieved and delivered in real time.

2. The method according to claim 1, characterized in that, The step of extracting knowledge from the customer's attribute feature data to obtain a basic customer knowledge graph includes: Data identification is performed on the customer's attribute feature data to distinguish between structured and unstructured data; For the structured data, customer attribute entities are extracted from it using preset mapping rules; For the unstructured data, named entity recognition technology is used to extract customer behavior entities, and relation extraction technology is used to identify the semantic association between the customer attribute entities and the customer behavior entities. The extracted customer attribute entities, customer behavior entities, and semantic relationships are organized into preliminary knowledge units in the form of triples. Based on the customer domain ontology model, knowledge fusion is performed on the preliminary knowledge units to eliminate entity ambiguity from different data sources and integrate the semantic relationships to generate integrated knowledge units. Based on the customer domain ontology model, define the entity types and relationship types of the basic customer knowledge graph, and populate the integrated knowledge units as instance data to construct the basic customer knowledge graph.

3. The method according to claim 1, characterized in that, The step of fusing the real-time consultation interaction data stream with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph includes: Streaming knowledge extraction is performed on the real-time consultation interaction data stream to extract intent entities, consultation focus, and contextual semantic relationships between the intent entities and consultation focus from the real-time consultation text and / or transcribed text of the consultation quote in real time, forming real-time intent knowledge units; Based on the cross-domain intent ontology model, the real-time intent knowledge unit is fused with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph. The cross-domain intent ontology model is used to define the semantic hierarchy and relational constraints between preset knowledge categories, and supports the establishment of entity links and semantic relationships between intent entities in the real-time intent knowledge unit and customer entities in the basic customer knowledge graph.

4. The method according to claim 3, characterized in that, The method based on a cross-domain intent ontology model integrates the real-time intent knowledge units with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph, including: The intent entity in the real-time intent knowledge unit is linked with the customer entity in the basic customer knowledge graph to identify the customer entity that points to the same real-world entity as the intent entity; the customer entity is the node representing the customer in the basic customer knowledge graph; the intent entity is the node representing the customer's consultation intent in the real-time consultation interaction data stream; the real-world entity is the customer that generates the real-time consultation interaction data stream. Based on the relational constraints defined by the cross-domain intent ontology model, a real-time semantic association is established between the intent entity and the identified customer entity, and the type of the real-time semantic association is determined by the relational constraints. The consultation focus and contextual semantic relationship associated with the intent entity are added to the customer intent knowledge graph as attribute information of the real-time semantic association, thus obtaining the dynamically updated customer intent knowledge graph.

5. The method according to claim 4, characterized in that, The step of linking the intent entities in the real-time intent knowledge unit with the customer entities in the basic customer knowledge graph includes: Extract the attribute features of the intent entity and the attribute features of the customer entity; Calculate the attribute similarity between the intent entity and each customer entity; The customer entity with the highest similarity exceeding a preset threshold is identified as the customer entity that points to the same real-world entity as the intent entity.

6. The method according to claim 1, characterized in that, The response to a change in the target customer's intent state in the customer intent knowledge graph satisfying a preset trigger condition, determining the target ad type based on the target customer's current intent tag, and retrieving the corresponding target ad for real-time delivery includes: Monitor the state changes of the intent nodes corresponding to the target customer in the customer intent knowledge graph. The state changes are triggered by knowledge processing operations performed on the customer intent knowledge graph. The knowledge processing operations include intent reasoning based on ontology rules and / or intent relationship completion based on graph embedding. When the change in the target customer's intent state is detected to meet the preset triggering conditions, the target customer's current intent tag is extracted; Based on the preset mapping relationship between intent tags and ad types, determine the target ad type corresponding to the current intent tag; Target ads that match the target ad type are retrieved from the ad resource library and delivered to the target customers in real time.

7. The method according to claim 6, characterized in that, The knowledge processing operations include: Obtain the intent entities, consultation focus, and real-time semantic associations associated with the target customer in the customer intent knowledge graph; Based on preset ontology rules, deductive reasoning is performed on the attributes of the intent entity and the real-time semantic associations to update the intent state of the target customer; and / or, The missing intent entities or real-time semantic associations in the customer intent knowledge graph are predicted using a knowledge graph embedding model. The customer intent knowledge graph is then completed based on the prediction results, and the intent state of the target customer is updated based on the completed graph.

8. The method according to claim 1, characterized in that, The preset triggering condition includes at least one of the following: The frequency of real-time inquiries from the target customer exceeded a preset threshold. The customer intent knowledge graph indicates that the risk intent level of the target customer has been upgraded from level one to level two; the first level is lower than the second level. The target customer made an inquiry about pricing within the preset time window.

9. A targeted advertising system based on a customer resource database, characterized in that, The system includes: The attribute data acquisition module is used to extract customer attribute feature data from the customer resource database in response to the login behavior triggered by the customer's access to the website; the customer attribute feature data includes the customer's login information and browsing information; The real-time data acquisition module is used to acquire real-time consultation interaction data stream in response to real-time consultation events. The real-time consultation interaction data stream includes at least real-time consultation text and / or consultation quote voice transcription text. The knowledge extraction module is used to extract knowledge from the attribute feature data of the customer to obtain a basic customer knowledge graph; wherein, the basic customer knowledge graph includes the semantic association between customer attributes and customer behavior; The knowledge fusion module is used to fuse the real-time consultation interaction data stream with the basic customer knowledge graph to obtain a dynamically updated customer intent knowledge graph. The advertising delivery module is used to respond to changes in the intent state of a target customer in the customer intent knowledge graph that meet preset trigger conditions, determine the target advertising type based on the target customer's current intent tag, and retrieve the corresponding target advertising for real-time delivery.

10. A targeted advertising platform based on a customer resource database, characterized in that, The platform includes: The professional knowledge service subsystem is used to provide professional knowledge training and quotation services, generate the real-time consultation events, and output the real-time consultation interaction data stream; The customer growth management subsystem is used to provide customer growth management services and to transmit customer conversion behavior data back. And the targeted advertising system as described in claim 9; The targeted advertising system establishes data connections with the professional knowledge service subsystem and the customer growth management subsystem to receive the real-time consultation interaction data stream and the customer conversion behavior data.

Images