Method and device for dynamically constructing personalized user portrait based on dialogue context

Abstract

This invention discloses a method and apparatus for dynamically constructing personalized user profiles based on dialogue context. It acquires multimodal dialogue streams and forms standardized text sequences; hierarchical context awareness obtains multi-level semantic features of users, forming a complete user attribute profile; structured instantaneous profile tags are generated and dynamically weighted; and multi-dimensional fusion and adaptive updating of user profiles are performed. The apparatus includes a dialogue interface and preprocessing module, a hierarchical semantic analysis engine, an instantaneous tag generation and weighting module, a profile fusion and update controller, a user profile database, and an asynchronous task scheduler. This invention transforms static, rigid user profiles into dynamic digital portraits that can reflect users' real-time states and autonomously evolve and update, achieving dynamic profile updates, accurate understanding of users' explicit and implicit needs, construction of an intelligent fusion mechanism for instantaneous and long-term profiles, and ensuring the timeliness and accuracy of profiles.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data processing technology, and relates to user profile construction and management technology. In particular, it relates to a method and device for dynamically constructing personalized user profiles based on dialogue context. Based on natural language processing, machine learning and dynamic weighted decay model technology, it realizes the dynamic construction and adaptive updating of personalized user profiles by analyzing the deep semantic information of human-computer dialogue flow in real time. Background Technology

[0002] In modern digital services, personalized recommendations and services have become core drivers for improving user experience and business efficiency. Whether it's e-commerce, content distribution, online advertising, or customer service, systems need to accurately understand each user's needs, preferences, and intentions. To achieve this "personalized experience for every individual," the industry widely adopts user profile technology, which analyzes various user data to assign a series of structured tags to each user, thereby creating a digital portrait of the user in the virtual world.

[0003] Especially in application scenarios where conversational AI is the core of interaction, such as intelligent customer service, virtual assistants, and online shopping guides, the dialogue interaction itself generates massive amounts of high-value unstructured text data. How to effectively utilize this dialogue data to build and optimize user profiles in order to provide more intelligent and personalized real-time responses is an important research direction and challenge in the current technology field.

[0004] Currently, there is a method for constructing static user profiles based on multi-source data aggregation. This method aggregates user-related information from multiple data sources, performs data cleaning, analysis, and statistics using offline batch processing, and transforms the extracted features into tags in a predefined tag system, which are then associated with the corresponding user identifiers. The system structure includes: (1) Data acquisition layer: Collects user data from different data sources, mainly including static attribute data provided by users during registration; user historical behavior data, such as browsing history and purchase history; and user-generated text content data, including comments and posts. (2) Data processing and analysis layer: Uses a batch processing framework to periodically process the collected data, and uses techniques such as keyword extraction, TF-IDF analysis, and topic modeling to identify user interests in the text data. (3) Profile storage layer: Stores the processed tag data in a database system and establishes a mapping relationship between users and tags. (4) Application interface layer: Provides user profile query interfaces for downstream application systems.

[0005] However, the existing technologies mentioned above have the following drawbacks: 1. Severe delay in profile updates: Due to the offline batch processing mode, the user profile update cycle is long, failing to reflect the real-time changes in user needs and intentions during the conversation, resulting in a significant lag between profile data and the actual user state. 2. Insufficient ability to understand dialogue context: Existing technologies mainly rely on keyword matching and statistical analysis methods when processing dialogue text, lacking a deep understanding of the context of multi-turn dialogues. They cannot accurately analyze linguistic phenomena such as pronouns and ellipsis, leading to inaccurate user intent recognition. 3. Lack of dynamic information fusion mechanism: Existing systems lack an effective mechanism to intelligently fuse new information generated during the conversation with existing user profiles, making it difficult to achieve dynamic updates and optimization of profiles. This leads to profile data gradually becoming rigid and unable to adapt to changes in user interests. 4. Incomplete interest evolution modeling: Existing technologies lack a scientific interest decay and update mechanism. Historical interest tags exist with fixed weights for a long time, failing to accurately reflect the natural evolution of user interests, affecting the accuracy and practicality of the profile.

[0006] Therefore, there is an urgent need for a new generation of user profile building technology that can deeply understand the context of dialogue, capture user dynamics in real time, and achieve intelligent integration and evolution of old and new information. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a method and apparatus for dynamically constructing personalized user profiles based on dialogue context. This addresses the core pain points of the prior art in terms of real-time user profiles, depth of contextual understanding, multi-dimensional information fusion, and timeliness management. It enables dynamic profile updates, accurate understanding of users' explicit and implicit needs, construction of an intelligent fusion mechanism for instantaneous and long-term profiles, and ensures the timeliness and accuracy of profiles.

[0008] The core of this invention lies in constructing a closed-loop system that integrates "layered semantic deep analysis" and "dynamic weighted intelligent evolution". By analyzing the dialogue context in real time, it extracts multi-dimensional features such as user intent, emotion, entity and attribute layer by layer, generates weighted instantaneous profile tags, and adopts dynamic weighted fusion and time decay mechanism to combine instantaneous insight with long-term profile intelligence and continuously evolve, thereby transforming the traditional static and rigid user profile into a dynamic digital portrait of the user that can reflect the user's real state in real time and evolve and update autonomously.

[0009] The technical solution provided by this invention is:

[0010] A method for dynamically constructing personalized user profiles based on dialogue context includes the following steps performed sequentially:

[0011] 1) Multimodal dialogue stream acquisition and normalization preprocessing;

[0012] Acquire multimodal dialogue data generated during user interaction with the application system platform, and standardize the raw data to form standardized text sequences. Specifically, this includes:

[0013] 1.1 Data Acquisition: Establish a real-time connection with the dialogue platform and continuously acquire multimodal dialogue streams, including text, speech transcription, and interactive events, through API interfaces via subscription or polling.

[0014] 1.2 Data Normalization: The acquired raw data is cleaned, colloquialized, segmented, part-of-speech tagged, stopped word filtered, and named entity linking is performed to form a standardized text sequence for subsequent analysis.

[0015] 2) Hierarchical context-aware feature extraction yields multi-level semantic features of the user, forming a complete user attribute profile;

[0016] This invention employs a multi-layered deep learning approach to perform hierarchical semantic deep parsing on preprocessed standardized text sequences, constructing a complete user feature system. In the feature extraction stage, this invention innovatively integrates multiple natural language processing models with a custom rule engine to form a hierarchical, context-aware parsing process, specifically including:

[0017] 2.1 Basic Semantic Parsing: A pre-trained named entity recognition model (e.g., BERT-NER) is used to accurately identify the entity objects (such as people) mentioned in the dialogue; and a relation extraction model (e.g., a relation classifier based on a pre-trained language model) is used to establish semantic associations between entities (such as "purchase-product"), thus constructing a basic semantic network.

[0018] 2.2 Intent and Sentiment Analysis: An intent classification model (e.g., a Transformer classifier) ​​is used to determine the user's core purpose in the current conversation round (e.g., "querying prices"); at the same time, an aspect-level sentiment analysis model (e.g., an attention-based LSTM model) is used to locate the user's sentiment tendency (positive, negative, neutral) and intensity towards the entity or aspect (e.g., represented by sentiment probability values ​​or intensity scores).

[0019] 2.3 Attribute Feature Extraction: This invention features specific construction in this step. First, a custom rule engine is incorporated, which incorporates grammatical and semantic rules (e.g., matching the sentence structure "I am [attribute]") to extract user attributes (such as "I live in Beijing") from explicit statements in the dialogue. Second, for implicit feature inference, a context-aware sequence labeling model is constructed. This model takes the outputs of steps 2.1 and 2.2 (entity, intent, sentiment) and the dialogue history as input, and annotates the user's implicit attributes (such as interests) by learning implicit cues in the dialogue context (e.g., the user's continued focus on "wooden products" implies their "preference"). Finally, the results of explicit extraction and implicit inference are combined to form a complete user attribute profile for the current dialogue turn.

[0020] 3) Instantaneous image tag structure generation and dynamic weighting;

[0021] 3.1 The multi-level semantic features extracted in step 2) are transformed into structured instantaneous profile tags with a unified format. Each tag includes metadata such as: tag content (e.g., "likes technology products"), category (e.g., "interest preferences"), polarity (e.g., positive), confidence score (probability value from the extraction model), and timestamp. Tags with confidence scores below a set threshold (e.g., 0.3) are automatically filtered and discarded to ensure that the data subsequently used in the fusion process has a reliable basic quality.

[0022] 3.2 Design a composite weighting function to calculate the initial weights of instantaneous image tags and evaluate the importance of instantaneous image tags;

[0023] To achieve a scientific assessment of tag importance, this invention designs a composite weighting function to calculate the initial weight W of each instantaneous image tag. initial This function comprehensively considers the following four dimensions of influence factors:

[0024] Explicitness E: Calculated based on the explicit / implicit method of feature extraction and the confidence level. A higher E value indicates a more explicit statement with higher confidence.

[0025] Emotional intensity S: The emotional intensity value is directly adopted from the aspect-level sentiment analysis output in step 2.2.

[0026] Positional Importance (P): Different weights are assigned based on the position of the relevant statements in the dialogue (such as the opening, the core discussion section, and the closing), with the core discussion section having a higher weight.

[0027] Recency effect R: Considering the proximity of the time when the label information appears to the current time, the more recent the information, the higher the R value.

[0028] The composite weighting function is specifically implemented in a linear weighted form:

[0029]

[0030] Here, β1, β2, β3, and β4 are adjustable weighting coefficients, summing to 1, used to balance the contributions of different dimensions. This function calculates the initial weights, ensuring the rationality and discriminative power of the weight allocation, laying the foundation for subsequent fusion and updates.

[0031] 4) Multi-dimensional fusion and adaptive updating of user profiles

[0032] Establish an intelligent fusion mechanism for instantaneous insights and long-term profiles to achieve continuous optimization and adaptive updates of user profiles.

[0033] 4.1 Semantic Alignment: Using vector similarity calculation (cosine similarity is used to calculate the similarity of tag embedding vectors), the instantaneous image tags are semantically matched with existing tags in the long-term image. A similarity threshold is set (usually ranging from 0.7 to 0.9). When the similarity exceeds this threshold, it is determined to be a successfully matched similar tag pair.

[0034] 4.2 Weight Fusion: For successfully matched tag pairs, a weighted moving average algorithm is used to fuse the weights of the old and new tags, updating the weight values ​​of the corresponding tags in the long-term profile. The update formula is:

[0035]

[0036] Among them, W new It is the new weight after merging. W instant This is the initial weight of the current instantaneous label. W old This refers to the old weight of the label in the long-term profile. α is a configurable dynamic learning rate. In this invention, α is not a fixed value, but is dynamically adjusted based on the label category (e.g., smaller α for more stable demographic attributes, larger α for more volatile interests and preferences) and the current dialogue scenario (e.g., features extracted in a complaint scenario may be more important), thereby balancing the contribution of new information and historical information.

[0037] 4.3 Knowledge Expansion: For instantaneous tags that fail to match but have a confidence level higher than a set threshold, they are added to the long-term profile as new profile dimensions, with their initial weight W. initial These initial values ​​enrich the user profile content. Other remaining tags are automatically filtered and discarded to avoid impacting the user profile data.

[0038] 4.4 Timeliness Maintenance: The tag weights in long-term profiles need to undergo time decay processing to ensure that the profile content maintains its timeliness and accuracy over time. For tags that have not been activated by new dialogue information recently, their weight W(t) decreases with time t (time since the last update).last_update It decreases as time goes on:

[0039]

[0040] Among them, W last_update This is the weight of the tag after the last update. λ is the decay coefficient. In this invention, λ can be set to different values ​​according to the tag category (for example, for stable attributes such as "occupation", λ is set to a smaller value or even 0; for volatile attributes such as "short-term interests", λ is set to a larger value) to achieve differentiated timeliness management and ensure that the profile content retains both memory and timeliness and accuracy in the long-term evolution.

[0041] The above four steps form a complete technical loop from data input to profile update. Through real-time dialogue analysis, deep semantic understanding, instant profile construction, and intelligent fusion update, the accurate construction and continuous evolution of user profiles are achieved.

[0042] In specific implementation, the present invention utilizes the above method to realize a personalized user profile dynamic construction device based on dialogue context, which can run on one or more servers and includes a processor and memory; the personalized user profile dynamic construction device includes:

[0043] • Dialogue Interface and Preprocessing Module: Configures multi-protocol adapters to connect to various dialogue platforms, has a built-in complete text preprocessing pipeline, and supports real-time data reception and normalization processing.

[0044] • Hierarchical Semantic Analysis Engine: Integrates multiple pre-trained language model instances, including dedicated sub-engines such as named entity recognition model, relation extraction model, intent classification model, and sentiment analysis model, to achieve hierarchical deep semantic understanding.

[0045] • Instantaneous label generation and weighting module: Generates structured instantaneous labels based on semantic analysis results, and assigns an initial weight value to each label through a dynamic weight calculation algorithm.

[0046] • Profile Fusion and Update Controller: Responsible for interacting with the user profile database, performing tag semantic matching, weight fusion calculation, and new tag admission control to ensure intelligent updates of profiles.

[0047] • User profile database: Adopts a hybrid storage solution, combining the advantages of NoSQL databases and graph databases, to support efficient storage of massive profile data and complex relationship queries.

[0048] • Asynchronous task scheduler: Manages the system's background computing tasks, including weight time decay calculation, model updates, etc., and ensures system performance and resource efficiency through task scheduling.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] This invention provides a method and apparatus for dynamically constructing personalized user profiles based on dialogue context. By constructing a closed-loop system integrating "layered semantic deep analysis" and "dynamic weighted intelligent evolution," it analyzes the dialogue context in real time, extracting multi-dimensional features such as user intent, emotion, entities, and attributes layer by layer to generate weighted instantaneous profile tags. A dynamic weighted fusion and time decay mechanism is employed to combine instantaneous insights with long-term profile intelligence and enable continuous evolution. This transforms traditional static and rigid user profiles into dynamic digital portraits that reflect the user's true state in real time and evolve autonomously. This invention addresses the core pain points of existing technologies in terms of real-time user profiles, depth of contextual understanding, multi-dimensional information fusion, and timeliness management. It achieves dynamic profile updates, accurate understanding of explicit and implicit user needs, construction of an intelligent fusion mechanism for instantaneous and long-term profiles, and ensures the timeliness and accuracy of the profiles. The specific technical advantages of this invention include:

[0051] 1. Significantly improved real-time performance and dynamism: This invention achieves dynamic profile updates through a real-time dialogue stream acquisition interface and an instantaneous tag generation mechanism. Compared with the traditional daily batch processing mode, it greatly improves the timeliness of profiles and can accurately capture the user's real-time intentions and status changes during the dialogue process.

[0052] 2. Significantly enhanced depth and accuracy of understanding: By employing a hierarchical deep learning model, including named entity recognition, relation extraction, intent classification, and aspect-level sentiment analysis, this invention breaks through the limitations of traditional keyword matching, significantly improves the accuracy of dialogue understanding, and is able to deeply understand contextual semantics, user sentiment, and complex intents.

[0053] 3. Outstanding intelligence and adaptability: This invention, through the organic combination of dynamic weighted calculation, intelligent fusion algorithm and time decay model, enables the system to have the ability to learn and optimize itself. It can automatically adjust the profile weight according to changes in user behavior and continuously maintain the accuracy and timeliness of the profile.

[0054] 4. Significantly improved scalability and practicality: The modular architecture design with loose coupling between functional components facilitates the expansion and maintenance of system functions, while supporting rapid deployment in multiple fields, significantly improving the practical value and applicability of the technology.

[0055] 5. Significantly improved information fusion efficiency: Through semantic similarity-based tag matching and intelligent fusion strategies, the problem of information silos in traditional systems is effectively solved, achieving efficient fusion of instantaneous and long-term profiles and forming a virtuous cycle of continuous optimization. Attached Figure Description

[0056] Figure 1 This is an overall flowchart of the method for dynamically constructing personalized user profiles based on dialogue context in an embodiment of the present invention.

[0057] Figure 2 This is a flowchart illustrating the internal structure and working process of the hierarchical semantic analysis engine in an embodiment of the present invention.

[0058] Figure 3 This is a block diagram showing the hardware deployment architecture of the user profile building system in this embodiment of the invention.

[0059] Figure 4 This is a block diagram of the composition structure of the personalized user profile dynamic construction device based on dialogue context in an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0061] The method and apparatus described in this invention constitute a complex system integrating real-time data processing, deep semantic analysis, dynamic weight management, and distributed storage. Its core lies in transforming raw dialogue data into continuously evolving personalized user profiles through an automated, continuously running closed-loop process. The successful implementation of this system relies on the collaborative design of the process, algorithm, and architecture, as detailed in the appendix. Figures 1 to 4 As shown, the implementation path of this solution is defined from four dimensions: logical flow, core engine, evolution mechanism, and physical deployment.

[0062] like Figure 1 As shown, the implementation of the method of the present invention includes the following four core steps:

[0063] 1. Multimodal dialogue stream acquisition and normalization preprocessing

[0064] This step establishes an efficient data acquisition and processing pipeline. The system receives real-time interactive data from multiple dialogue platforms, including text dialogues, speech transcriptions, and structured interaction events, through a distributed API gateway cluster deployed at the access layer. In the data preprocessing stage, the system executes a multi-level processing flow: first, deep text cleaning removes irrelevant symbols and noise; second, grammatical correction and expression standardization are performed using a neural network-based language model; then, domain-adaptive word segmentation technology, combined with a professional dictionary, achieves accurate word segmentation, obtaining entities to build a high-quality data foundation for subsequent in-depth analysis.

[0065] 2. Hierarchical context-aware feature extraction, from bottom to top, includes: basic semantic layer, intent and sentiment layer, and user attribute and fact layer, to obtain user attribute profile;

[0066] This step employs a three-layer progressive architecture to achieve deep semantic understanding. The basic semantic layer, based on a BERT-BiLSTM-CRF hybrid model, accurately identifies named entities in the dialogue and constructs an entity relationship network. The intent and sentiment layer utilizes a multi-task learning framework to simultaneously classify user intent and perform aspect-level sentiment analysis, accurately capturing the purpose and sentiment of the dialogue. The user attribute and fact layer comprehensively utilizes a custom rule engine and a context-aware sequence labeling model to infer the user's deep characteristics from explicit statements and implicit context. The three-layer architecture transmits data through well-defined structured data objects (such as JSON Schema), with each layer adhering to predetermined input and output specifications. The output of the lower layer provides contextual support for the analysis of the upper layer, ensuring the completeness and accuracy of semantic understanding.

[0067] 3. Instantaneous portrait tag structure generation and dynamic weighting

[0068] This step transforms semantic features into standardized profile data units. Based on the output of step 2, the system constructs structured instantaneous tags containing metadata such as tag content, category system, and confidence level. The system establishes a quality verification mechanism, automatically filtering and discarding tags with a confidence level below a set threshold (e.g., 0.3) to ensure the accuracy and reliability of the profile data; these tags do not participate in weight calculations. An innovative composite weight function calculates a scientifically reasonable initial weight for each tag. The calculation process is as follows: the system receives the extracted confidence level C and the original sentiment intensity value S of the tag. raw The system identifies the number of dialogue rounds (m), the total number of dialogue rounds (N), and the extraction method. First, it quantifies four dimensions. The explicitness of expression (E) is determined by the extraction method, with a coefficient (k) set to 1.0 for explicit statements and 0.7 for implicit inferences. Then, it is calculated as E = k * C. The emotional intensity (S) is set to S... raw The absolute value of the positional importance P is assigned a preset value based on the interval of the m / N ratio (e.g., 1.0 for the core segment, 0.8 for the opening segment, and 0.6 for the ending segment). The recency effect R is calculated as R = m / N. Subsequently, the system substitutes the four dimension values ​​into the linear weighting formula W. initial The formula W is calculated as β1*E + β2*S + β3*P + β4*R, where the weighting coefficients (β1, β2, β3, β4) are preset values ​​(e.g., 0.4, 0.3, 0.2, 0.1). The final calculated W is... initial It is written into the tag metadata.

[0069] 4. Multi-dimensional fusion and adaptive updating of user profiles

[0070] This step enables continuous optimization and evolution of user profile data. The system maintains a long-term user profile database, which is initialized during the user's first interaction with the system and continuously evolves in all subsequent conversations. The system first uses semantic vector similarity calculation to intelligently match the instantaneous tag set generated in the current conversation with existing tags in the long-term profile database. For successfully matched tags, an adaptive weighted moving average algorithm is used for weight fusion, where the learning rate is dynamically adjusted based on tag category and business scenario. Unmatched high-confidence tags are incorporated into the profile system through a strict admission mechanism. Other tags are automatically filtered and discarded to avoid affecting the fusion and establishment of user profiles. Simultaneously, the system performs timely maintenance on long-term profiles, reducing the weight of tags that have not been activated by new conversation information over time to ensure that the profiles maintain both freshness and stable features.

[0071] In specific implementation, the present invention utilizes the above method to realize a personalized user profile dynamic construction device based on dialogue context, which can run on one or more servers and includes a processor and memory; the personalized user profile dynamic construction device includes:

[0072] • Dialogue Interface and Preprocessing Module: Configures multi-protocol adapters to connect to various dialogue platforms, has a built-in complete text preprocessing pipeline, and supports real-time data reception and normalization processing.

[0073] • Hierarchical Semantic Analysis Engine: Integrates multiple pre-trained language model instances, including dedicated sub-engines for named entity recognition, relation extraction, intent classification, and sentiment analysis, to achieve hierarchical deep semantic understanding.

[0074] like Figure 2 As shown, the hierarchical semantic analysis engine employs a fine-tuned architecture based on a pre-trained language model to construct a hierarchical semantic understanding system. This engine adopts a three-layer progressive design: the basic semantic layer serves as the underlying support, constructing a basic semantic framework through named entity recognition and relation extraction; the intent and sentiment layer serves as the intermediate processing unit, simultaneously performing intent classification and aspect-level sentiment analysis to accurately capture the user's dialogue purpose and emotional inclination; and the user attribute and fact layer serves as the top-level reasoning module, inferring the user's deep attribute characteristics from explicit statements and implicit context. Each sub-engine communicates efficiently through standardized interfaces, forming a complete semantic understanding chain, ensuring both the depth of analysis and improving the system's maintainability.

[0075] • Instantaneous label generation and weighting module: Generates structured instantaneous labels based on semantic analysis results, and assigns initial weight values ​​to each label through a composite weight function.

[0076] • Profile Fusion and Update Controller: Responsible for interacting with the user profile database, performing tag semantic matching, weight fusion calculation, and new tag admission control to ensure intelligent updates of profiles.

[0077] • User profile database: Adopts a hybrid storage solution, combining the advantages of NoSQL databases and graph databases, to support efficient storage of massive profile data and complex relationship queries.

[0078] • Asynchronous task scheduler: Manages the system's background computing tasks, including weight time decay calculation, model updates, etc., and ensures system performance and resource efficiency through task scheduling.

[0079] like Figure 3 As shown, this invention employs a microservice-based architecture deployed in a containerized environment and orchestrated using Kubernetes to build a highly available user profiling service platform. The access layer ensures high service availability through load balancing and an API gateway; the computing layer adopts a dual-pipeline design, combining real-time and asynchronous processing to guarantee rapid response to online requests while efficiently processing batch tasks through message queues; the data layer uses a hybrid storage scheme based on data type, combining the advantages of NoSQL databases, graph databases, and caching systems to achieve efficient data access. The entire architecture, through resource isolation and elastic scaling, ensures low-latency response while possessing excellent scalability and fault tolerance, capable of supporting large-scale concurrent user business scenarios.

[0080] like Figure 4 As shown, this invention provides a personalized user profile dynamic construction device based on dialogue context. Through the hierarchical collaboration and asynchronous scheduling of six modules, it achieves real-time dynamic updates of the profile. First, the dialogue interface and preprocessing module receive dialogue data from multiple platforms, and after normalization, output text data to the hierarchical semantic analysis engine. This engine performs deep semantic parsing through multiple pre-trained model sub-engines, passing the results to the instantaneous label generation and weighting module to generate structured labels with initial weights. Subsequently, the profile fusion and update controller receives the label data, interacts bidirectionally with the user profile database, completes label matching, weight fusion, and valid label admission, and writes the updated profile data to the database. Simultaneously, the asynchronous task scheduler independently sends instructions to each core module: a model update instruction to the hierarchical semantic analysis engine, a weight time decay calculation instruction to the instantaneous label generation and weighting module, a fusion strategy adjustment instruction to the profile fusion and update controller, and a storage optimization scheduling instruction to the user profile database. This ensures efficient operation of the main process while optimizing system performance and resources, ultimately achieving accurate iteration of the user profile.

[0081] Through the above steps, this invention forms a complete technical process: real-time dialogue stream input, hierarchical semantic parsing, instantaneous profile generation, dynamic fusion and updating, profile service output, and continuous autonomous evolution. This process not only ensures the real-time nature and accuracy of profile updates, but also endows the system with self-optimization capabilities through a closed-loop feedback mechanism, possessing high intelligence and universal applicability.

[0082] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A method for dynamically constructing personalized user profiles based on dialogue context, characterized in that, Includes the following steps: 1) Acquire the multimodal dialogue stream and perform normalization preprocessing to form a standardized text sequence; 2) Hierarchical context-aware feature extraction yields multi-level semantic features of the user, forming a complete user attribute profile; The natural language processing model is collaboratively built with a custom rule engine to form a hierarchical, context-aware parsing process; including: 2.1 Basic Semantic Parsing: Identify entity objects in the dialogue, establish semantic relationships between entities, and construct a basic semantic network; 2.2 Intent and Sentiment Analysis: Determine the user's core purpose in the current conversation round; Identify the user's emotional inclination and intensity towards entities or aspects; 2.3 Attribute Feature Extraction; Extraction methods include: First, a custom rule engine is built-in to extract syntactic and semantic features of user attributes from explicit statements in the dialogue. Second, for implicit feature inference, a context-aware sequence labeling model is constructed, using entities, intentions, sentiments, and dialogue history as inputs to the sequence labeling model. By learning implicit cues in the dialogue context, the implicit attributes of users are labeled. Finally, the results of explicit extraction and implicit inference are combined to form a user attribute profile. 3) Instantaneous image tag structure generation and dynamic weighting; 3.1 The extracted multi-level semantic features of users are transformed into structured instant profile tags with a unified format; if the confidence level of the structured instant profile tags is lower than the set threshold, they are automatically filtered and discarded. 3.2 Design a composite weighting function to calculate the initial weights of instantaneous profile tags and evaluate the importance of the instantaneous profile tags; the composite weighting function is expressed as: in, 1 represents the initial weight of the instantaneous profile label; E represents the clarity of expression; S represents the emotional intensity; P represents the positional importance; R represents the recency effect; β1, β2, β3, and β4 are adjustable weighting coefficients that sum to 1, used to balance the contributions of different dimensions. 4) Establish an intelligent fusion mechanism for instantaneous insights and long-term profiles, enabling multi-dimensional fusion and adaptive updates of user profiles; including: 4.1 Perform semantic alignment by semantically matching the instantaneous profile tags with existing tags in the long-term profile to identify successfully matched similar tag pairs; 4.2 Perform weight fusion: For successfully matched tag pairs, fuse the weights of the old and new tags, and update the weight values ​​of the corresponding tags in the long-term profile; the update method is expressed as follows: Among them, W new It is the new weight after fusion; W instant It is the initial weight of the current instantaneous label; W old It is the old weight of this label in the long-term profile; It is a configurable dynamic learning rate; 4.3 Knowledge Expansion: For instantaneous tags that fail to match but have a confidence level higher than a set threshold, add them as new profile dimensions to the long-term profile, using their initial weight W. initial As initial value; 4.4 Timeliness Maintenance: The weights of tags in long-term profiles undergo time decay to ensure the timeliness and accuracy of the profile content over time. For tags that have not been activated by new dialogue information recently, their weight W(t) decreases as time t increases from the last update time; expressed as: Where W(t) is the label weight; W last_update This represents the weight of the label after the last update; λ is the decay coefficient; t is time; t last_update This is the last update time.

2. The method for dynamically constructing personalized user profiles based on dialogue context as described in claim 1, characterized in that, Step 2.1 Perform basic semantic parsing, specifically by using a pre-trained named entity recognition model to identify entity objects in the dialogue; and by using a relation extraction model to establish semantic relationships between entities.

3. The method for dynamically constructing personalized user profiles based on dialogue context as described in claim 2, characterized in that, Step 2.2 Perform intent and sentiment analysis. Specifically, use an intent classification model to determine the user's core purpose in the current conversation round. At the same time, use an aspect-level sentiment analysis model to locate the user's sentiment tendency and intensity towards entities or aspects. Sentiment tendency includes positive, negative, or neutral. Sentiment intensity is represented by sentiment probability value or intensity score.

4. The method for dynamically constructing personalized user profiles based on dialogue context as described in claim 3, characterized in that, In step 3.1, the structured instantaneous profile tags with a unified format include: tag content, category, polarity, confidence level, and timestamp.

5. The method for dynamically constructing personalized user profiles based on dialogue context as described in claim 4, characterized in that, In step 3.2, the explicitness of expression is specifically calculated based on the explicit / implicit method of feature extraction and the confidence level. Explicit statements with high confidence levels result in a high explicitness value. The sentiment intensity is specifically calculated using the sentiment intensity value output by aspect-level sentiment analysis. The positional importance is assigned different weights based on the position of the tag-related statements in the dialogue. The recency effect is specifically derived from the proximity of the time when the label information appears to the current moment.

6. The method for dynamically constructing personalized user profiles based on dialogue context as described in claim 4, characterized in that, In step 4.1, the cosine similarity is used to calculate the similarity of the label embedding vectors; and a similarity threshold is set to determine the successfully matched similar label pairs.

7. A device for dynamically constructing personalized user profiles based on dialogue context using the method described in claim 1, characterized in that, The device includes a processor and memory, and runs on one or more servers; The device comprises the following modules: a dialogue interface and preprocessing module, a hierarchical semantic analysis engine, an instantaneous tag generation and weighting module, a profile fusion and update controller, a user profile database, and an asynchronous task scheduler.

8. The personalized user profile dynamic construction device based on dialogue context as described in claim 7, characterized in that, The dialogue interface and preprocessing module are used to configure multi-protocol adapters to connect to various dialogue platforms. It has a built-in text preprocessing pipeline and supports real-time data reception and normalization processing. The hierarchical semantic analysis engine is used to integrate multiple pre-trained language model instances to achieve hierarchical deep semantic understanding; The instantaneous label generation and weighting module is used to generate structured instantaneous labels based on semantic analysis results, and assign initial weight values ​​to the labels by calculating dynamic weights; The profile fusion and update controller is used to intelligently update profiles by interacting with the user profile database to perform tag semantic matching, weight fusion calculation, and new tag admission control. The user profile database uses a hybrid storage solution of NoSQL database and graph database to support efficient storage of massive profile data and complex relationship queries. An asynchronous task scheduler is used to manage the system's background computing tasks, ensuring system performance and resource efficiency through task scheduling.

9. The personalized user profile dynamic construction device based on dialogue context as described in claim 8, characterized in that, The hierarchical semantic analysis engine specifically includes a named entity recognition model, a relation extraction model, an intent classification model, and a sentiment analysis model.

Images