A cross-cultural dynamic portrait-driven multi-agent content generation method and system

Abstract

This invention discloses a multi-agent content generation method and system driven by cross-cultural dynamic profiling, belonging to the fields of artificial intelligence and cross-cultural information processing technology. The method includes: S1, collecting multilingual text data through social media interfaces and performing structural verification, language identification, text normalization, deduplication, and quality filtering to construct a cross-cultural corpus knowledge base; S2, dividing the corpus into groups based on clustering algorithms, extracting group features, and constructing structured user profiles; S3, scheduling multiple agents to collaboratively perform content generation, trend analysis, and cross-cultural interaction tasks according to user profile tags and rule-based routing; S4, constructing a virtual audience cluster and pre-evaluating the generation strategy using a simulation mechanism that decouples the generation model and the evaluation model; S5, optimizing and iterating the generation strategy based on the evaluation results to form a closed-loop update mechanism. This invention effectively improves the stability and computational accuracy of the cross-language data processing link.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, big data processing and computer software technology, and specifically relates to a method for constructing dynamic user profiles, multi-agent collaborative decision-making and content generation for cross-cultural communication scenarios. Background Technology

[0002] Existing content generation and decision support technologies for cross-cultural communication scenarios typically rely on large language models, rule engines, or static tagging systems to generate communication content, interactive language, and strategic suggestions. However, due to significant differences in language habits, cultural contexts, and audience psychology among different target markets, the adaptability of existing technologies in complex cross-cultural communication scenarios still faces the following challenges.

[0003] On the one hand, raw data from multilingual social media platforms typically exhibits characteristics such as high noise, strong heterogeneity, and structural instability. Existing systems often struggle to balance data cleaning accuracy and computational overhead when processing massive cross-lingual corpora, resulting in insufficient semantic purity of the underlying data. Simultaneously, existing user profiling methods largely rely on static rules, manually preset labels, or fixed clustering methods, making it difficult to adaptively track dynamic changes in the distribution of target market data. Furthermore, it is challenging to further transform clustering results into standardized data structures that can be directly used by downstream agents, thus affecting the accuracy of profiling results and the stability of the system chain.

[0004] On the other hand, existing technologies mostly employ serial processing or single-model generation chains in the strategy generation and execution stages, lacking adaptive routing and multi-agent collaborative mechanisms to address audience sentiment changes, cultural taboos, and market differences. Furthermore, existing solutions are typically unidirectional outputs, lacking simulation pre-evaluation methods based on target cultural characteristics before formal strategy deployment. This makes it difficult to identify potential risks such as cultural friction, semantic misinterpretation, and insufficient content adaptation in a timely manner, thus hindering the formation of a closed-loop optimization mechanism for generation, evaluation, and correction.

[0005] Therefore, how to improve the quality of cross-cultural multilingual corpus processing, realize multi-agent collaborative decision-making driven by dynamic profiling, and establish a closed loop for strategy evaluation and feedback optimization oriented towards target cultural scenarios has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the technical challenges of existing technologies in cross-cultural communication scenarios, such as high noise levels, unstable structure, static user profiles, insufficient multi-agent collaborative capabilities, and lack of a closed-loop strategy evaluation and feedback optimization mechanism in multilingual social media data, this invention provides a cross-cultural dynamic profile-driven multi-agent content generation method, computer equipment, and computer-readable storage medium. This aims to improve the quality of cross-cultural multilingual corpus processing, achieve multi-agent collaborative decision-making under dynamic profile-driven conditions, and establish a closed-loop strategy evaluation and feedback optimization mechanism oriented towards the target cultural scenario.

[0007] To address the above problems, the present invention provides the following technical solution: A method for generating multi-agent content driven by cross-cultural dynamic profiling includes the following steps: S1: Collect multilingual social media data streams through data interfaces, and perform structure verification, language identification, text normalization, noise removal, deduplication, desensitization processing, and feature extraction on the data streams to obtain structured corpus data; S2: Cluster the semantic features, sentiment features, interaction features, and market features of the structured corpus data, determine the target number of clusters based on the clustering quality assessment results, and extract representative texts from each cluster; input the representative texts into the large language model to generate a dynamic user profile that meets the preset data structure constraints; S3: Using the tag information in the dynamic user profile as the conditional routing input, multiple heterogeneous agents are scheduled to perform serial and / or parallel processing, and the output results of each heterogeneous agent are aggregated to obtain the structured strategy result. S4: Construct a weighted virtual audience cluster based on the market characteristic distribution of the dynamic user profile, input the structured strategy results into the weighted virtual audience cluster for multiple rounds of simulation interaction, and generate strategy evaluation results based on the simulation interaction results; S5: Correct the structured strategy result based on the risk identifier in the strategy evaluation result, and perform a second simulation interaction based on the corrected structured strategy result; write the evaluation result obtained from the second simulation interaction into the knowledge base, and update the feature vector and profile label weight through automated update tasks to form an iterative optimization closed loop.

[0008] Preferably, step S1 includes: performing key field integrity verification on the multilingual social media data stream, retaining text fields, time fields, language identifier fields, interaction fields, source fields, and market identifier fields; performing language identification and splitting based on the language identifiers and language detection results inherent in the data stream; performing encoding unification, character standardization, abnormal whitespace folding, noise information removal, and topic word splitting on the text; performing a first-stage duplicate content filtering based on rule matching; generating text semantic vectors using a sentence vector encoding model, and performing a second-stage semantic duplicate content filtering based on the cosine similarity between any two text semantic vectors; when the cosine similarity is greater than a preset numerical threshold, the corresponding text is determined to be semantically duplicated text; and performing desensitization processing and feature extraction on the processed samples to obtain the structured corpus data.

[0009] Preferably, step S2 includes: constructing a clustering input vector based on semantic features, sentiment features, interaction features, and market features; performing clustering operations on multiple candidate clusters respectively, and determining the target number of clusters based on the clustering quality assessment results; extracting representative text for each cluster; and inputting the clustering results and the representative text into a large language model to generate a dynamic user profile that meets preset data structure constraints.

[0010] Preferably, the clustering operation uses K-means clustering, the clustering quality assessment result is the silhouette coefficient, and the target number of clusters is determined based on the average of the silhouette coefficients of all sample points; the dynamic user profile includes at least user demand tags, media preference tags, cultural constraint tags, technology preference tags, and target market tags.

[0011] Preferably, step S3 includes: parsing the tag information, sentiment information, and market information in the dynamic user profile to generate a conditional routing control signal; triggering multiple heterogeneous agents to perform serial processing and / or parallel processing according to the conditional routing control signal; and aggregating the output results of the multiple heterogeneous agents according to a unified field structure to obtain the structured strategy result.

[0012] Preferably, the plurality of heterogeneous intelligent agents include at least an audience insight and content generation intelligent agent, a product optimization and service upgrade intelligent agent, a trend analysis intelligent agent, a cross-cultural interaction intelligent agent, and a virtual brand testing intelligent agent; wherein, the parallel processing of the audience insight and content generation intelligent agent and the trend analysis intelligent agent is triggered based on technology preference tags; the product problem attribution processing of the product optimization and service upgrade intelligent agent is triggered based on the proportion of negative emotions in the emotional information exceeding a preset threshold; the secondary verification processing of the cross-cultural interaction intelligent agent is triggered based on the target market belonging to a high-context culture; and the pre-evaluation processing of the virtual brand testing intelligent agent is triggered based on the result of generating a new strategy.

[0013] Preferably, steps S4 and S5 include: performing stratified sampling based on the market characteristic distribution and profile weights of the dynamic user profile to construct a weighted virtual audience cluster; configuring user attribute parameters for each virtual audience agent in the weighted virtual audience cluster; driving the virtual audience agent to simulate its interactive behavior in the target cultural context, and scoring the content effect and cultural risk corresponding to the interactive behavior based on preset rules; weighting and summarizing the results of multiple rounds of simulated interaction to generate the strategy evaluation result; correcting the structured strategy result based on the risk identifier in the strategy evaluation result, and inputting the corrected structured strategy result into the weighted virtual audience cluster to perform another simulated interaction; and writing the evaluation result obtained from the second simulated interaction into the knowledge base to update the feature vector and profile tag weights.

[0014] Preferably, the method further includes an output standardization step, which includes: establishing preset data structure constraints for dynamic user profiles, structured strategy results, and strategy evaluation results; performing field integrity checks, field type checks, and constraint condition checks before each object enters the downstream processing node; triggering a retry when the check fails; and rolling back to the most recently checked object version when the retry still fails, and recording the corresponding audit log.

[0015] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the cross-cultural dynamic portrait-driven multi-agent content generation method as described above.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the cross-cultural dynamic portrait-driven multi-agent content generation method described above are implemented.

[0017] Compared with the prior art, the present invention has one or more of the following beneficial technical effects: This invention improves the quality of corpus processing in high-noise and highly heterogeneous data environments by performing structural verification, rule filtering, and semantic duplicate content filtering on multilingual social media data streams, providing a more stable data foundation for subsequent dynamic user profile generation.

[0018] This invention combines semantic features, emotional features, interaction features, and market features for clustering, and dynamically determines the target number of clusters based on the clustering quality assessment results. This enables audience segmentation to better adapt to changes in current data distribution, thereby improving the accuracy and usability of dynamic user profiles.

[0019] This invention uses tag information from dynamic user profiles as conditional routing input to schedule multiple heterogeneous intelligent agents to perform serial and / or parallel processing, thereby achieving coordinated execution of processing steps such as content generation, trend analysis, product problem attribution, cross-cultural verification, and simulation evaluation, thus improving the relevance of strategy output and processing efficiency.

[0020] This invention, by constructing a weighted virtual audience cluster and conducting multiple rounds of simulation interaction before formal output, can identify risks such as cultural misinterpretation, insufficient content adaptation, and comprehension barriers before actual delivery, thereby reducing the trial-and-error costs in cross-cultural communication.

[0021] This invention improves the system's dynamic adaptability to changes in the target market and the dissemination environment by writing simulation evaluation results into a knowledge base and combining automated update operations to update feature vectors and profile label weights. Attached Figure Description

[0022] Figure 1 This is the main flowchart of the cross-cultural dynamic profiling-driven multi-agent content generation method provided in this embodiment of the invention, and is also included as an abstract figure.

[0023] Figure 2 This is a hierarchical architecture diagram for data processing and dynamic user profile generation provided in the embodiments of the present invention.

[0024] Figure 3 This is a diagram of the multi-agent conditional routing and cooperative execution architecture provided in an embodiment of the present invention.

[0025] Figure 4 This is a closed-loop diagram of simulation evaluation and feedback optimization provided in the embodiments of the present invention.

[0026] Figure 5 This is a structural block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art to the technical solutions of the present invention without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0028] To address the problems of high noise levels, structural instability, static user profiles, insufficient multi-agent collaborative capabilities, and lack of simulation evaluation and feedback optimization loops in existing cross-cultural communication content generation technologies, this invention proposes a cross-cultural dynamic profile-driven multi-agent content generation method and system. This method uses multilingual social media data as input and, through data preprocessing, dynamic user profile generation, multi-agent conditional routing and collaborative execution, virtual audience simulation evaluation, and feedback optimization updates, forms a closed-loop content generation chain for cross-cultural communication scenarios. This improves the quality of structured corpus processing, the targeting of strategy generation, and the dynamic adaptability of communication schemes.

[0029] Figure 1 The main flowchart of the cross-cultural dynamic profile-driven multi-agent content generation method provided in this embodiment of the invention is shown. Figure 2 This diagram illustrates the layered architecture for data processing and dynamic user profile generation provided in an embodiment of the present invention. Figure 3 This diagram illustrates a multi-agent conditional routing and cooperative execution architecture provided in an embodiment of the present invention. Figure 4 The simulation evaluation and feedback optimization closed-loop diagram provided in the embodiment of the present invention is shown; Figure 5 A block diagram of a computer device structure provided in an embodiment of the present invention is shown.

[0030] like Figure 1 As shown, the cross-cultural dynamic profile-driven multi-agent content generation method provided in this embodiment of the invention includes the following steps S1 to S5.

[0031] S1: Collect multilingual social media data streams and process them to obtain structured corpus data. like Figure 1 and Figure 2 As shown, step S1 is used to acquire multilingual social media data streams in cross-cultural communication scenarios and convert them into structured corpus data that can be used for subsequent profile generation and multi-agent processing. Specifically, the system receives multilingual text data streams through a social media data interface. These multilingual text data streams include at least text fields, time fields, language identifier fields, interaction fields, source fields, and market identifier fields. Input data first enters... Figure 2 The preprocessing layer shown.

[0032] In this embodiment, the preprocessing layer includes a structure verification unit, a language identification and splitting unit, a text normalization unit, a noise stripping and quality filtering unit, a rule-based deduplication unit, a semantic deduplication unit, and a desensitization unit. Specifically, the structure verification unit checks the completeness and format validity of data fields, removing or recording data samples with missing key fields, abnormal field types, or those that cannot be parsed; the language identification and splitting unit performs dual language identification based on the language identifiers inherent in the data stream and the language detection results, and splits samples of different languages ​​to their corresponding processing queues; the text normalization unit performs encoding unification, character normalization, abnormal whitespace folding, case unification, and topic word splitting; the noise stripping and quality filtering unit removes invalid links, meaningless duplicate characters, advertising template fragments, and low-information-density text; the rule-based deduplication unit identifies explicit duplicate content based on rule matching; the semantic deduplication unit identifies semantically duplicate content; and the desensitization unit performs desensitization or anonymization processing on sensitive information such as user identifiers, contact information, and precise location.

[0033] In semantic deduplication, preferably, a sentence vector encoding model is used to map the text into semantic vectors, and the cosine similarity between any two candidate texts is calculated. The calculation formula is as follows:

[0034] in, Indicates the first The semantic vector of the text. Indicates the first The semantic vector of the text. Indicates the semantic similarity between two texts. When When the value exceeds a preset threshold, the corresponding text is identified as semantically repetitive text, and texts with higher information completeness, richer interactive information, or more recent time are retained as representative samples.

[0035] After preprocessing is complete, the data enters... Figure 2 The feature construction layer is shown. This layer includes semantic feature extraction units, sentiment feature extraction units, interaction feature extraction units, and market feature extraction units, which are used to extract feature information of their respective dimensions and form the input features required for subsequent clustering and profile generation. Finally, the processed results form structured corpus data and are output to... Figure 2 The output layer shown is simultaneously written to a persistent storage unit for subsequent profile generation, playback auditing, and knowledge retrieval.

[0036] S2: Generate dynamic user profiles based on multidimensional feature clustering and extracting representative text. In step S2, to avoid algorithmic black box effects, the dynamic user profile generated by the large language model adopts a strict JSON data structure contract. The system mandates that this JSON object must be predefined and contain the following core fields: appeal_tag, tech_pref, target market_tag, negative_emotion_ratio, sentiment_info, and market_info. The system configures a structured input prompt template for the large language model, with the input feature format strictly defined as {cluster_id, top_texts, sentiment_ratio, market_code}. Simultaneously, the temperature parameter in the decoding strategy is strictly locked at 0.2 to reduce the divergence and feature drift rate of the output results. At the output node, the system independently deploys an integrity constraint engine to perform strong validation of the returned JSON fields for field integrity and type. If the verification fails, the system triggers an automatic repair retry based on the error log. If the retry fails twice in a row, the current abnormal fluctuation result is discarded, and the object version rollback mechanism is triggered to retrieve the stable profile version that passed the verification in the previous historical time window, thereby ensuring that the downstream processing link does not crash or stop due to data structure mismatch.

[0037] like Figure 1 and Figure 2 As shown, step S2 is used to generate a dynamic user profile based on the structured corpus data obtained in step S1. Specifically, the semantic features, sentiment features, interaction features, and market features from the structured corpus data are input... Figure 2 The image generation layer shown in the diagram performs clustering operations on the input samples by the clustering processing unit, and evaluates the quality of the clustering results corresponding to different candidate cluster numbers by the clustering quality assessment unit, thereby determining the target number of clusters.

[0038] In this embodiment, preferably, the clustering processing unit uses the K-means clustering algorithm to perform clustering; the clustering quality evaluation unit preferably uses the silhouette coefficient as the clustering quality index. The corresponding silhouette coefficient calculation formula is:

[0039] in, Represents the total number of samples. Indicates the first The average distance between each sample and samples in the same cluster Indicates the first The average distance between each sample and its nearest neighbor cluster sample Represents the number of candidate clusters The corresponding average silhouette coefficient. The system calculates the average silhouette coefficient based on the number of candidate clusters. The target number of clusters is selected by choosing the candidate clusters with better silhouette coefficients.

[0040] After determining the target number of clusters, a representative text extraction unit extracts representative texts for each cluster. Preferably, samples within a cluster can be sorted according to factors such as the distance between the sample and the cluster center, the intensity of text interaction, and the recentity of the sample, and the top-ranked texts are selected as representative texts. Subsequently, the representative texts are input into a large language model tag generation unit, which generates a dynamic user profile that meets preset data structure constraints. The dynamic user profile includes at least user demand tags, media preference tags, cultural constraint tags, technology preference tags, and target market tags. In some embodiments, it may also include fields such as negative sentiment percentage, high-context cultural markers, suggested expressions, and prohibited expressions.

[0041] To ensure that dynamic user profiles can be stably used by subsequent multi-agent systems. Figure 2 The profile structure verification unit shown is used to perform field integrity verification, field type verification, and constraint condition verification on the profile object output by the large language model. When the verification fails, a retry is triggered; if the retry still fails, the process rolls back to the most recently verified profile version. Through the above processing, a structured and directly callable dynamic user profile is obtained and written to the persistent storage unit.

[0042] S3: Using profile tags as conditional routing input, schedule heterogeneous intelligent agents for collaborative processing. like Figure 1 and Figure 3 As shown, step S3 is used to schedule multiple heterogeneous agents to perform collaborative processing based on the tag information in the dynamic user profile, and obtain a structured policy result. Specifically, the dynamic user profile is first input... Figure 3 The conditional routing control node shown parses the tag information, sentiment information, and market information in the profile and generates conditional routing control signals according to preset routing rules.

[0043] In this embodiment, the heterogeneous intelligent agents include at least an audience insight and content generation intelligent agent, a product optimization and service upgrade intelligent agent, a trend analysis intelligent agent, a virtual brand testing intelligent agent, and a cross-cultural interaction intelligent agent. The conditional routing control node triggers the corresponding intelligent agent to participate in processing based on different tags or state conditions. For example, when a technology preference tag is detected, the audience insight and content generation intelligent agent and the trend analysis intelligent agent are triggered; when the proportion of negative sentiment exceeds a preset threshold, the product optimization and service upgrade intelligent agent is triggered; when a new strategy is detected, the virtual brand testing intelligent agent is triggered; and when the target market is identified as a high-context culture, the cross-cultural interaction intelligent agent is triggered to perform cultural adaptation verification. Figure 3 Each agent can perform serial and / or parallel processing based on routing relationships, and finally summarize the output results into a structured policy result.

[0044] Among them, the Audience Insight and Content Generation agent generates content candidates, expression style suggestions, and media adaptation solutions based on dynamic user profiles; the Product Optimization and Service Upgrade agent generates product problem attribution and improvement suggestions for samples with concentrated negative sentiment; the Trend Analysis agent identifies hot topic changes and competitor trends; the Virtual Brand Testing agent performs pre-evaluation before official output; and the Cross-Cultural Interaction agent performs expression risk verification, cultural taboo identification, and localization adjustment suggestions based on the target market's cultural context. The final input of the content output by each agent is... Figure 3 The bottom structured strategy results aggregation node forms a structured strategy result with a unified field structure.

[0045] In a preferred embodiment, the trend analysis agent can identify potential trend breakout points by observing changes in the proportion of topics within a continuous time window. For a given topic within a given time window... The percentage is recorded as Then the standardized formula for the intensity of change can be expressed as:

[0046] in, Indicates the rate of change of the proportion of topics over time. This represents the mean of the rate of change within the corresponding time window. This represents the standard deviation of the rate of change within the corresponding time window. This indicates the strength of the change after standardization. When... When the number of windows exceeds a preset threshold and continues to reach a preset number, the topic can be identified as a potential trend breakout point.

[0047] S4: Construct a weighted virtual audience cluster and execute multiple rounds of simulation interaction to generate evaluation results. like Figure 1 and Figure 4As shown, step S4 is used to pre-evaluate the results of the structured strategy before the actual deployment. Specifically, firstly, a weighted virtual audience cluster is constructed based on the market characteristic distribution and profile weights of the dynamic user profile, and user attribute parameters are configured for each virtual audience agent. The user attribute parameters may include one or more of age group, price sensitivity, technology preference, media preference, cultural sensitivity, and brand familiarity. Subsequently, the results of the structured strategy are input into the weighted virtual audience cluster, and multiple rounds of simulation interaction are performed to generate evaluation results.

[0048] In this embodiment, Figure 4 The phrase "inputting structured strategy results and executing multi-round simulation interactions" corresponds to the process of delivering the structured strategy results to a virtual audience agent and collecting its simulated feedback. Preferably, the simulation interaction employs a mechanism that decouples the generative model from the evaluation model, where the generative model simulates the interactive behavior of the virtual audience agent in the target cultural context, and the evaluation model scores the simulation results. Scoring dimensions may include one or more of content effectiveness, cultural risk, comprehension difficulty, and acceptance tendency. After weighted summarization of the multi-round simulation interaction results, the strategy evaluation result is obtained.

[0049] S5: Modify the strategy based on the risk identification and simulate again, then execute feedback updates to form a closed loop. like Figure 1 and Figure 4 As shown, step S5 is used to correct the structured strategy result based on the risk identifier in the strategy evaluation result, and write the corrected result back to the knowledge base to form a feedback optimization closed loop. Specifically, when Figure 4 When a judgment node determines that a risk indicator exists, it corrects the structured strategy result to obtain a corrected structured strategy result; then, the corrected structured strategy result is input again into the simulation interaction link to perform a second simulation. This process corresponds to... Figure 4 The "correcting the structured strategy result" in the middle flows back to the closed-loop path of "inputting the structured strategy result and performing multiple rounds of simulation interaction".

[0050] In this embodiment, the risk markers may include risks of cultural misinterpretation, inappropriate expression, comprehension barriers, emotional triggering, and other communication risks. Different correction methods can be used for different risk markers. For example, for the risk of cultural misinterpretation, high-risk cultural symbols can be replaced, festival mappings can be adjusted, or language style can be adjusted; for the risk of comprehension barriers, explanatory content can be added, terminology density can be reduced, or a description method more in line with local expression habits can be adopted; for the risk of inadequate content adaptation, the title, media format, call to action, or narrative focus can be adjusted.

[0051] when Figure 4When a judgment node determines that no risk identifier exists, the evaluation result is written to the knowledge base and a feedback update is performed. The feedback update includes at least feature vector updates and profile label weight updates. By writing the evaluation result from a re-simulation or final pass to the knowledge base, and combining this with automated update operations to incrementally update the feature vector and profile label weights, an iterative optimization closed loop adapted to changes in the target market and the communication environment can be formed.

[0052] In a preferred embodiment, the present invention further includes an output standardization step. This output standardization step establishes unified data structure constraints for dynamic user profiles, structured policy results, and policy evaluation results, and performs field integrity checks, field type checks, and constraint condition checks before each object enters the downstream processing node. If a check fails, a retry is triggered; if the retry also fails, the process rolls back to the most recently validated object version and records the corresponding audit log. This output standardization step improves the consistency of data interfaces between nodes in a multi-agent collaborative link and reduces link anomalies caused by structural inconsistencies.

[0053] This invention also provides a multi-agent content generation system driven by cross-cultural dynamic profiling. The system is used to execute the method steps described in the above embodiments. (Combined with...) Figures 2 to 4 The system includes at least a data processing and dynamic user profile generation subsystem, a multi-agent conditional routing and cooperative execution subsystem, a simulation evaluation and feedback optimization subsystem, and a persistent storage subsystem.

[0054] The data processing and dynamic user profile generation subsystem is used to execute steps S1 and S2, and includes at least an input layer, a preprocessing layer, a feature construction layer, a profile generation layer, and an output layer; the multi-agent conditional routing and cooperative execution subsystem is used to execute step S3, and includes at least a conditional routing control node and multiple heterogeneous agents; the simulation evaluation and feedback optimization subsystem is used to execute steps S4 and S5, and includes at least a weighted virtual audience cluster construction unit, a multi-round simulation interaction unit, an evaluation result generation unit, a risk identification judgment unit, and a feedback update unit; the persistent storage subsystem is used to store structured corpus data, dynamic user profiles, structured policy results, policy evaluation results, feature vectors, and profile label weights.

[0055] The system can be deployed centrally or in a distributed manner. Preferably, the subsystems interact with each other through message passing, interface calls, shared storage, or a combination thereof. Any system configuration that adopts the above-described intelligent agent division and can implement the technical solution of this invention should fall within the protection scope of this invention.

[0056] like Figure 5As shown, the present invention also provides a computer device. The computer device includes at least a processor, a memory, a system bus, a network interface, and an input / output interface. The processor, the memory, the network interface, and the input / output interface all interact with each other via the system bus. The memory stores a computer program, which, when executed by the processor, implements the steps of the cross-cultural dynamic portrait-driven multi-agent content generation method described in the above embodiments.

[0057] Preferably, the processor can be a central processing unit, a graphics processing unit, an artificial intelligence acceleration processing unit, or other processing unit capable of executing instructions; the memory can include read-only memory, random access memory, flash memory, solid-state drive, or other non-volatile and / or volatile storage media; the network interface is used to communicate with external social media interfaces, knowledge bases, logging systems, and other external devices; the input / output interface is used to receive manual configuration, output profile results, output strategy results, or output evaluation results.

[0058] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the cross-cultural dynamic profile-driven multi-agent content generation method described in the above embodiments. The computer program may include one or more of the following: a multilingual data acquisition and preprocessing program, a dynamic user profile generation program, a multi-agent conditional routing program, a simulation evaluation program, and a feedback update program. Any computer-readable storage medium that employs the above program logic and can implement the technical solution of this invention should fall within the protection scope of this invention.

[0059] Those skilled in the art should understand that the data processing system provided by this invention strictly adheres to data security and privacy compliance requirements and does not violate public interests. The data interfaces used in step S1 are all officially authorized public social media API interfaces. Before constructing text features and performing clustering calculations, the system's desensitization agent enforces a structured cleaning procedure, physically truncating and deleting all sensitive personal privacy data such as user identifiers (ID / username), contact information, and precise geographical location. Simultaneously, the dynamic user profile tag set configured in the system backend is strictly constrained by a compliance whitelist, and its algorithm logic pre-excludes sensitive tag features such as gender, ethnicity, and religious beliefs that involve discrimination or violate public order and good morals through regular expression filtering. Before the evaluation results and features are finally written into the system knowledge base, data fairness verification and compliance auditing must be performed through a front-end security gateway to ensure that the technical solution of this invention fully complies with current laws and regulations while improving the underlying information processing efficiency of the computer system.

Claims

1. A multi-agent content generation method driven by cross-cultural dynamic profiling, characterized in that, Includes the following steps: S1: Collect multilingual social media data streams through data interfaces, and perform structure verification, language identification, text normalization, noise removal, deduplication, desensitization processing, and feature extraction on the data streams to obtain structured corpus data; S2: Cluster the semantic features, sentiment features, interaction features, and market features of the structured corpus data, determine the target number of clusters based on the clustering quality assessment results, and extract representative texts from each cluster; input the representative texts into the large language model to generate a dynamic user profile that meets the preset data structure constraints; S3: Using the tag information in the dynamic user profile as the conditional routing input, multiple heterogeneous agents are scheduled to perform serial and / or parallel processing, and the output results of each heterogeneous agent are aggregated to obtain the structured strategy result. S4: Construct a weighted virtual audience cluster based on the market characteristic distribution of the dynamic user profile, input the structured strategy results into the weighted virtual audience cluster for multiple rounds of simulation interaction, and generate strategy evaluation results based on the simulation interaction results; S5: Correct the structured strategy result based on the risk identifier in the strategy evaluation result, and perform another simulation interaction based on the corrected structured strategy result; The evaluation results obtained from the re-simulation interaction are written into the knowledge base, and the feature vector and portrait label weights are updated through automated update tasks to form an iterative optimization closed loop.

2. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 1, characterized in that, Step S1 includes: Perform key field integrity checks on the multilingual social media data stream, retaining text fields, time fields, language identifier fields, interaction fields, source fields, and market identifier fields; Language identification and stream segmentation are performed based on the language identifiers and language detection results inherent in the data stream. The text is subjected to encoding standardization, character normalization, abnormal whitespace folding, noise information removal, and topic word splitting. The first stage of duplicate content filtering is performed based on rule matching. Text semantic vectors are generated using a sentence vector encoding model, and then based on any two text semantic vectors... and cosine similarity Perform the second stage of semantic duplicate content filtering; when the cosine similarity is greater than a preset numerical threshold, the corresponding text is determined to be semantic duplicate text; The processed samples are subjected to desensitization and feature extraction to obtain the structured corpus data.

3. The cross-cultural dynamic profiling-driven multi-agent content generation method according to claim 1, characterized in that, Step S2 includes: Clustering input vectors are constructed based on semantic features, sentiment features, interaction features, and market features; Clustering operations are performed on multiple candidate cluster numbers, and the target number of clusters is determined based on the clustering quality assessment results. Extract representative text from each cluster; The clustering results and the representative text are input into a large language model to generate a dynamic user profile that meets the preset data structure constraints.

4. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 3, characterized in that: The clustering operation uses K-means clustering; The clustering quality assessment result is the silhouette coefficient, which is based on the average distance between a sample point and the samples within the same cluster. and the average distance to the nearest neighbor cluster samples Calculated as The target cluster number is determined based on the average of the contour coefficients of all sample points; The dynamic user profile includes at least user demand tags, media preference tags, cultural constraint tags, technology preference tags, and target market tags.

5. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 1, characterized in that, Step S3 includes: Analyze the tag information, sentiment information, and market information in the dynamic user profile to generate conditional routing control signals; The conditional routing control signal triggers multiple heterogeneous agents to perform serial and / or parallel processing. The outputs of the multiple heterogeneous agents are aggregated according to a unified field structure to obtain the structured strategy result.

6. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 5, characterized in that, The multiple heterogeneous intelligent agents include at least an audience insight and content generation intelligent agent, a product optimization and service upgrade intelligent agent, a trend analysis intelligent agent, a cross-cultural interaction intelligent agent, and a virtual brand testing intelligent agent. Among them, the parallel processing of audience insight and content generation intelligent agents and trend analysis intelligent agents is triggered based on technology preference tags; Based on the fact that the proportion of negative emotions in the emotional information exceeds a preset threshold, the defect attribution processing of the product optimization and service upgrade intelligent agent is triggered. Secondary verification processing of cross-cultural interactive intelligent agents is triggered based on the target market belonging to a high-context culture. The pre-evaluation process of the virtual brand test agent is triggered based on the results of generating new strategies.

7. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 1, characterized in that, Steps S4 and S5 include: Based on the market characteristic distribution and profile weight of the dynamic user profile, stratified sampling is performed to construct a weighted virtual audience cluster; Configure user attribute parameters for each virtual audience agent in the weighted virtual audience cluster; The virtual audience agent is driven to simulate its interactive behavior in the target cultural context, and the content effect and cultural risk corresponding to the interactive behavior are scored based on preset rules. The results of multiple rounds of simulation interaction are weighted and summarized to generate the strategy evaluation result; The structured strategy results are corrected based on the risk indicators in the strategy evaluation results, and the corrected structured strategy results are then fed into the weighted virtual audience cluster to perform another simulation interaction. The evaluation results obtained from the re-simulation interaction are written into the knowledge base to update the feature vector and profile label weights.

8. The multi-agent content generation method driven by cross-cultural dynamic profiling according to claim 1, characterized in that, The method further includes an output standardization step, which includes: Establish preset data structure constraints for dynamic user profiles, structured strategy results, and strategy evaluation results; Before each object enters the downstream processing node, perform field integrity checks, field type checks, and constraint checks. Trigger a retry if the verification fails; If the retry still fails, roll back to the most recently validated object version and record the corresponding audit log.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the cross-cultural dynamic portrait-driven multi-agent content generation method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the cross-cultural dynamic portrait-driven multi-agent content generation method as described in any one of claims 1 to 8.

Images