A digital human virtual human setup generation method

Abstract

The application discloses a digital person virtual person setting generation method, and relates to the technical field of data processing. The method comprises the following steps: in response to an input task theme, a target region identifier and an industry classification identifier, generating a person setting genotype structure according to the target region identifier and the industry classification identifier; constructing a generation environment control file based on the person setting genotype structure; loading the generation environment control file into a pre-trained generation model to generate differentiated content according to the task theme; analyzing the cognitive homogenization risk of the differentiated content through a graph isomorphism detection algorithm; and if the graph similarity of the cognitive homogenization risk exceeds a preset threshold, updating the person setting genotype structure. The application reduces the homogenization degree of the content output by a large model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for generating digital human virtual personas. Background Technology

[0002] The booming development of social e-commerce overseas has created immense value for high-quality social media accounts. Among these, "persona" is a key element in shaping an account's image and a core aspect of building a robust operational environment. A clearly defined and distinctive persona can effectively increase account traffic, platform ranking, and recommendation rates.

[0003] The large model can output the required persona materials in batches based on the basic positioning of the marketing audience. It can also generate images, daily posts, voice messages and short videos on a regular basis according to the persona, providing high-quality content for enterprises to operate overseas personas.

[0004] However, the content output by large models suffers from cross-account homogenization. For example, when the same algorithm is used to manage 100 3C review accounts from different countries: a Brazilian account posts "In-depth teardown report on iPhone heat dissipation performance", a German account posts "iPhone heat dissipation structure dissection experiment", and a Japanese account posts "iPhone heat dissipation system decomposition research". On the surface, the content is compliant, but in reality, the core is homogenized, which makes it easy to be punished by the platform's algorithm and reduce the account's weight.

[0005] The homogenization of the core content makes each piece of content compliant when viewed individually, and the persona parameters do not deviate. All key indicators monitored by existing technology (such as sensitive words and interaction volume) are normal. However, since operators have to manage hundreds of accounts, it is impossible to compare each piece of content one by one. This leads to the account weight declining silently, and by the time it is discovered, the account is often already paralyzed.

[0006] Therefore, how to reduce the homogeneity of the output content of large models has become a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The technical problem solved by this invention is that the content output by large models has a high degree of homogeneity across accounts.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for generating a digital human virtual persona, comprising: responding to an input task theme, target region identifier, and industry classification identifier; generating a persona genotype structure based on the target region identifier and industry classification identifier; constructing a generation environment control file based on the persona genotype structure; loading the generation environment control file into a pre-trained generation model; generating differentiated content based on the task theme; analyzing the cognitive homogenization risk of the differentiated content using a graph isomorphism detection algorithm; and updating the persona genotype structure if the graph similarity of the cognitive homogenization risk exceeds a preset threshold.

[0009] Preferably, in response to the input task theme, target region identifier, and industry classification identifier, a persona genotype structure is generated based on the target region identifier and industry classification identifier. This includes: in response to the input task theme, target region identifier, and industry classification identifier, calling CLDR to extract the regional feature dataset corresponding to the target region identifier; retrieving the target knowledge graph corresponding to the industry classification identifier from the pre-constructed industry knowledge graph based on the industry classification identifier; performing pruning operations on the target knowledge graph based on the importance of nodes in the target knowledge graph to delete non-core entity nodes and retain core entity nodes, forming an industry-specific entity set; and storing the regional feature dataset and the industry-specific entity set into the persona genotype structure.

[0010] Preferably, based on the persona genotype structure, a generation environment control file is constructed, including: determining the information entropy of the industry-specific entity set; configuring a dynamic sampling temperature function that fluctuates with the generation time period according to the information entropy; statistically analyzing the occurrence frequency of sequence chains composed of multiple entities within the industry-specific entity set; identifying entity sequence chains exceeding the frequency threshold as high-frequency logical chain patterns to be suppressed; constructing an attention mask matrix based on the high-frequency logical chain patterns, the attention mask matrix being used to suppress word sequences matching the high-frequency logical chain patterns; and constructing the generation environment control file based on the dynamic sampling temperature function and the attention mask matrix.

[0011] Preferably, the generation environment control file is loaded into the pre-trained generative model, and differentiated content is generated according to the task theme, including: obtaining the original corpus according to the task theme; using an abstract semantic representation parser to decompose the original corpus into semantic elements to obtain a set of semantic elements; assigning weights to the semantic elements in the set of semantic elements that are associated with specific regional terms according to the usage frequency information of specific regional terms; arranging the semantic elements in the set of semantic elements according to the assigned weight values ​​to obtain a sequence of semantic elements; inputting the sequence of semantic elements into the generative model, and controlling the randomness of the generation process of the generative model through a dynamic sampling temperature function to output differentiated content.

[0012] Preferably, the cognitive homogenization risk of differentiated content is analyzed by a graph isomorphism detection algorithm, including: representing differentiated content as a labeled directed graph consisting of semantic role nodes and logical relationship edges between nodes; determining the similarity between the labeled directed graph corresponding to the differentiated content and the labeled directed graph corresponding to the historically generated content based on a preset graph isomorphism detection algorithm; and determining the cognitive homogenization risk of differentiated content based on the similarity.

[0013] Preferably, if the graph similarity of the cognitive homogenization risk exceeds a preset threshold, the persona genotype structure is updated, including: if the similarity corresponding to the cognitive homogenization risk exceeds the preset threshold, the historical usage frequency of each entity node in the industry-specific entity set stored in the persona genotype structure within the cross-account scope is statistically analyzed to obtain an overuse rate index; a threshold value is determined based on information entropy and cognitive homogenization risk; entity nodes in the industry-specific entity set whose overuse rate index exceeds the threshold value are removed; new entity nodes are selected from the pre-built cross-domain knowledge base and injected into the industry-specific entity set to complete the update of the persona genotype structure.

[0014] Preferably, the semantic elements associated with specific regional terms in the semantic element set are weighted according to the usage frequency information of specific regional terms, including: determining the breadth of regional coverage based on the number of administrative regions covered by the industry-specific entity set and the usage frequency information of specific regional terms; determining the semantic relevance of multiple words in the original corpus to the industry-specific entity set; determining the conflict index of multiple words in the original corpus based on the semantic relevance; and determining the importance of each word among the multiple words based on the breadth of regional coverage and the conflict index.

[0015] Preferably, after determining the importance of each word among multiple words based on the geographical coverage and conflict index, the method further includes: sequentially inputting the original corpus into a religiously sensitive word database and a regional folklore word database for filtering; performing a substitution word mapping transformation on the filtered words to obtain the transformed semantic element set and the importance of the words in the semantic element set.

[0016] Preferably, determining the information entropy of the industry-specific entity set includes: determining the entity distribution status of the industry-specific entity set: statistically analyzing the proportion of different types of entities in the industry-specific entity set; determining the information entropy of the industry-specific entity set based on the proportion of different types of entities, wherein a higher entropy value indicates a more discrete category distribution of entities in the industry-specific entity set, and a higher degree of dispersion in the category distribution of entities in the industry-specific entity set indicates stronger industry knowledge diversity.

[0017] Preferably, the dynamic sampling temperature function that fluctuates with the generation time period is configured according to the information entropy, including: adjusting the fluctuation amplitude coefficient and base value parameter of the preset temperature function according to the entropy value of the information entropy to obtain the dynamic sampling temperature function, wherein the fluctuation amplitude coefficient is directly proportional to the information entropy value, and the base value parameter is inversely proportional to the information entropy value.

[0018] The beneficial effects of this invention are as follows: By generating a persona genotype structure, based on target regional identifiers and industry classification identifiers, it accurately captures regional cultural context and industry knowledge boundaries, thereby eliminating redundant information and initial persona ambiguity caused by traditional static prompts, thus laying a fundamental foundation for differentiated content generation; Based on the persona genotype structure, a generation environment control file is constructed, and by systematically adjusting the key parameters of the generation model, it avoids the style rigidity and content template-based approach caused by existing technologies relying on one-time prompts, dynamically guiding the content generation process towards deep localization adaptation; After loading the generation environment control file into the pre-trained generation model, differentiated content is generated according to the task theme, by deconstructing thematically mixed information and incorporating regional characteristics... From a substantive perspective, this approach optimizes content in both clarity of focus and depth of regional adaptation. It analyzes the cognitive homogenization risk of generated content using a graph isomorphism detection algorithm, automatically identifying core homogenization that surface indicators cannot capture by using the similarity of the content's internal logical structure graph as an indicator. This reveals hidden risks at the essential level. When the graph similarity exceeds a preset threshold, it triggers an update to the persona genotype structure, correcting genotype deviations at the source and forming a closed-loop feedback mechanism. This prevents the accumulation of homogenization patterns in subsequent content, systematically improving the uniqueness and localization quality of the content. It avoids platform algorithms being penalized due to homogenization, ultimately achieving efficient generation of highly differentiated and credible virtual persona content, enhancing user experience and content dissemination. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the basic process of a digital human virtual persona generation method provided in one embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for generating a digital human virtual persona is provided, comprising:

[0022] S110, responding to the input task theme, target region identifier, and industry classification identifier, generates a persona genotype structure based on the target region identifier and industry classification identifier.

[0023] S120, based on the human character genotype structure, constructs and generates environment control files.

[0024] S130 loads the generated environment control file into the pre-trained generative model and generates differentiated content based on the task theme.

[0025] S140 uses a graph isomorphism detection algorithm to analyze the risk of cognitive homogenization of differentiated content.

[0026] S150: If the graph similarity of the cognitive homogenization risk exceeds the preset threshold, then update the persona genotype structure.

[0027] If the graph similarity of the perceived risk of homogenization does not exceed the preset threshold, then the differentiated content generated in S130 will be published.

[0028] Specifically, S110 generates a persona genotype structure in response to the target region identifier and industry classification identifier input by the user.

[0029] Traditional content generation relies on static, general prompts (such as "Please write a review about iPhone heat dissipation, aimed at Brazilian tech enthusiasts"). This approach ignores the unique cultural context of the target region (e.g., Brazilian users are more concerned with actual performance under high outdoor temperatures), the core and boundaries of industry knowledge (in 3C reviews, "heat dissipation" is a basic item, while "energy efficiency optimization" may have greater potential for differentiation), and the necessity of avoiding redundant information. This results in a vague initial "persona" for the generated content, sowing the seeds for subsequent homogenization.

[0030] Current technologies cannot systematically digitize the precise characteristics of regions and industries and eliminate irrelevant noise. Therefore, when a user inputs a target region identifier (e.g., BR for Brazil) and an industry category identifier (e.g., Consumer_Electronics_Review for consumer electronics reviews) through a front-end interface or API, the server receives the input and initializes a structured data container (called a "structure"). This structure is the core basis for all subsequent steps; it is not a simple collection of tags, but a multi-dimensional vector or data object containing parameters that define the essence of the persona.

[0031] Subsequently, CLDR (Unicode CommonLocaleDataRepository) was invoked to extract the regional feature dataset. CLDR is an international standard database that provides data on formatting rules, symbols, and naming conventions for global languages, regions, and cultures. CLDR was applied here to accurately capture Brazil's "localized" characteristics. CLDR not only provides language translation rules but also includes typical name formats and common first / last names, date, time, and number representation conventions (such as the currency symbol R, thousands / decimal point separator conventions), regionally specific holidays or cultural events, and localized vocabulary preferences for specific topics (such as technology, sports, and weather) (not simple translations, but commonly used local expressions). In practice, the server executes the following queries: extracts the date format of Brazil (BR) as dd / MM / yyyy (different from the MM / dd / yyyy of the United States), extracts the Brazilian Real symbol and its format rule as R1.234,56, extracts a list of common Brazilian male / female names (for reference on cultural authenticity when generating virtual character names later), queries common Brazilian Portuguese words related to "high temperature" and "outdoors" and their frequencies (such as calorintenso - intense heat, praia - beach, carnaval - carnival, as related words), and obtains data on major Brazilian cities (such as Rio de Janeiro and São Paulo) and their typical summer temperature ranges (if CLDR is extended or associated with meteorological data). The final extracted geographic feature dataset is a collection of key-value pairs (e.g., a JSON structure), for example: {"locale":"pt_BR","currency":{"symbol":"R$","format":"#.##0,00"},"dateFormat":"dd / MM / yyyy","commonTopics":{"heat":["calor intenso"," quente"],"outdoor":["praia","parque"]},"cityTemps":{"Rio deJaneiro":[25,40]},...}.

[0032] Next, non-core entity nodes are removed using an industry-specific knowledge graph pruning algorithm, retaining only the industry-specific entity set. General or industry-specific knowledge graphs contain a vast number of entities (such as iPhone, heat sink, CPU, battery, design aesthetics, brand value, Apple Inc., Samsung Inc.) and their relationships. For a specific domain (such as consumer electronics reviews), not all entities are equally important. Core entities refer to the elements that directly constitute the core content of the review (functions, performance metrics, comparison objects). Non-core entities (such as Apple Inc.'s financial reports, the influence of art styles in design aesthetics) may distract the model, introduce information noise unrelated to the "review" theme, or cause the generated content to deviate from the core, increasing the risk of unintentional homogenization (e.g., content all touching on company strategy).

[0033] The pruning algorithm aims to focus on the core of the industry and remove unnecessary details. Its implementation includes: inputting a SPARQL query: SELECT ? entity WHERE { ? entity rdf:type:Consumer_Electronics_Review. ? entity:hasPageRank? prValue} ORDER BY DESC(?prValue) LIMIT 100. This is a standard knowledge graph query language (SPARQL). Here, ? entity rdf:type:Consumer_Electronics_Review means finding all entities of type:Consumer_Electronics_Review (assuming such a category exists in the knowledge graph, representing the entities involved in the evaluation); ? entity:hasPageRank?prValue requires these entities to have a PageRank value attribute (:hasPageRank); ORDER BY DESC(?prValue) sorts them by PageRank value from highest to lowest; LIMIT 100 only retrieves the top 100 most important entities after sorting (100 is the preset number of core entities N). The PageRank algorithm (derived from web search) is used to evaluate the importance of nodes (entities) in a knowledge graph. An entity receives more links / references from other important entities, resulting in a higher PageRank. In a benchmark graph, the iPhone 15 Pro's thermal performance is likely to be linked to by multiple cooling solutions (graphene, vapor chamber) and comparison models (Samsung S24, Pixel 8 Pro), leading to a high PageRank. Apple's Q4 2023 financial report, on the other hand, may only have a few benchmark links (primarily involving price analysis), resulting in a lower PageRank.

[0034] To address the issue that static thresholds cannot adapt to the varying sparsity of entity graphs across different industries (a mature industry like mobile phone reviews has dense entity relationships and many entities with high PageRank values; a new industry like AR glasses reviews may have sparse entity relationships), gradient descent is used to optimize the threshold (dynamically): an initial threshold (e.g., 0.1) is set; the algorithm monitors the "usefulness" of the pruned core entity set in historical content generation (e.g., the frequency with which the core entities are used by the model, or the performance of the generated content in subsequent homogenization detection); if the relevance of the content generated by the core entity set decreases or remains highly homogenized, the PageRank threshold is slightly reduced using gradient descent (e.g., from 0.1 to 0.08) to include slightly more, but potentially slightly marginal, core entities; conversely, the threshold is slightly increased. The goal is to find a threshold that allows the core entity set to cover the core points without being too large and causing distraction.

[0035] The pruning operation includes: after executing the query, not only retaining the top 100 entities (part of the core entity set), but also traversing all entities belonging to that industry category; for any leaf node with a PageRank value lower than the current (possibly optimized) threshold (i.e., the end node in the knowledge graph with an out-degree of 0, no longer connected to other important information), such as Steve Jobs' design philosophy (if it is an isolated, low PageRank node in the graph), it is removed from the core considerations. After filtering and sorting, the core entity list of the Brazilian consumer electronics review account is obtained: ['Heat dissipation performance', 'Battery life', 'Processor speed', 'Screen display', 'Camera quality', 'Gaming performance', 'Drop test', 'Samsung Galaxy S24', 'Google Pixel 8', 'Outdoor use', 'High temperature environment', 'Value for money', 'Local warranty'...]. Note: The region identifier BR will subtly affect this list. For example, Brazil's high temperatures may indirectly affect or subsequently adjust the weights of 'Outdoor use' and 'High temperature environment' based on CLDR data.

[0036] CLDR precisely anchors regional characteristics, preventing all accounts from uniformly describing the "test environment," laying a solid foundation for injecting localized differentiation (for example, German accounts might use CLDR DE to extract related terms such as engineering precision and rigorous standards); focusing on the core and eliminating interference sources—industry knowledge graph pruning clearly defines the boundaries of "what to evaluate," eliminating non-core interference items, preventing the model from deviating from the topic or all accounts from falling into generalities (such as all talking about company history), forcing content to find differentiation points within the defined core areas; dynamic optimization—gradient descent optimization of the PageRank threshold ensures that the core entity set can adapt to the structure of different industry graphs and the constantly evolving information environment, maintaining its accuracy and effectiveness. The advantages of this solution compared to existing technologies are reflected in the following aspects: It goes beyond simple domain tags—existing technologies may only add locale=BR to prompts or provide a translation dictionary. This solution provides structured, in-depth cultural and habitual data through CLDR, achieving true localization integration; it goes beyond keyword libraries and fixed templates—traditional methods rely on manually maintained keyword libraries or templates. This solution, based on industry knowledge graphs and graph algorithms (PageRank), automatically, quantitatively, and dynamically identifies and retains core entities, making it more intelligent, scalable, and adaptable. Manual maintenance cannot handle complex relationships and dynamic changes.

[0037] S120, Based on the aforementioned human genotype structure, construct and generate an environment control file.

[0038] Existing large-scale model content generation primarily relies on user prompts for one-time control, lacking a systematic and dynamic adjustment mechanism for key parameters during the generation process. This leads to: rigid style and creativity—the model tends to generate "safe and average" content. For example, when generating technical evaluations, the model may default to a flat, descriptive style, piling up technical parameters, resulting in all accounts having similar expressions and perspectives (e.g., all resembling excerpts from instruction manuals); and a tendency towards content templates—the model easily recognizes and uses frequently occurring logical structures (e.g., "test objective → test method → ​​data display → test conclusion"), meaning that even content from different accounts may have highly isomorphic internal logical chains. In this implementation, the input is the persona genotype structure generated in the previous step (containing the localeFeatures_BR dataset and the industry-specific entity set coreEntities), and the output is a structured control file (e.g., JSON or YAML) that guides the model's internal generation behavior.

[0039] Configure the dynamic sampling temperature function: T(t) = T0 + k·sin(2π·t / Tc). The sampling temperature T is a core hyperparameter controlling the randomness and creativity of the large model's output. When T = 0, the model always chooses the word with the highest probability (the output is deterministic, conservative, and predictable); as T increases, the model is more inclined to choose words with lower probabilities (the output is more random, diverse, and creative, but may also be incoherent). Fixing T will lead to a single generation style: if the temperature is too low, the content will be rigid and inflexible; if the temperature is too high, the content may deviate from the topic or lose its professionalism. Function Explanation: T0 (base value) is the default base temperature value, determined by the character genotype. For example, for a German account with high professional rigor (this requirement may be implicit in the genotype), T0 might be set to 0.7; for an entertainment and science popularization account emphasizing creativity and fun (this requirement may be implicit in the genotype), T0 might be set to 1.2; k (innovation coefficient) is the amplitude of temperature fluctuation, determined by the character genotype. For example, if the genotype is marked "high creativity index", then k is set to 0.4; if it is marked "stability priority", then k is set to 0.1 (k = 0 is equivalent to a fixed temperature); Tc (cycle control) determines the sine wave cycle (i.e., the "frequency" of temperature change). For example, Tc = 50 (tokens) controls the rhythm of creative fluctuations; t (generation time) represents the t-th token generated by the current model. The function effect is that the sampled temperature fluctuates periodically between T0±k: when the T value is high (peak), the model output is more creative and may choose novel metaphors, unusual perspectives or expressions; when the T value is low (trough), the model output is more stable and professional, ensuring the accuracy of core information and data representation.

[0040] Specific example: For a Brazilian 3C account (genotype set as T0=1.0, k=0.3, Tc=60), when the model generates an article describing the product's appearance and initial impression (requiring some creativity), t is between 0-15 (at or close to Tmax≈1.3), resulting in a more vivid output (e.g., "The smooth titanium frame of the iPhone 15 Pro, held under the Rio sun, feels like holding a delicate, slightly warm beach pebble..."); when entering the core performance test data presentation section (requiring precision), t is between 30-45 (at Tmin≈0.7), resulting in a more rigorous output. Dynamic temperature simulation mirrors the thought process of a human author—sometimes divergent and creative, sometimes rigorous and convergent—breaking the monotony of the generated content's rhythm and expression, achieving a balance between content fluency and stylistic diversity, and increasing the micro-differences between different content under the same persona.

[0041] An attention mask matrix is ​​generated, applying zero weights to a predefined sequence of homogeneous logical chains. The large model is based on the Transformer architecture, whose core mechanism is self-attention. When predicting the next word, the model calculates the association strength (i.e., attention weights) between the current word and all previous words (called context). These weights determine the degree of influence of the context on the generation of the current word. The attention mask matrix is ​​a matrix (or vector) of the same length as the input token sequence, and its element values ​​are used to adjust the original attention weights (usually through multiplication). Assigning zero weights means forcibly masking certain specific contextual information. In specific domains, models can form frequently used fixed logical patterns. For example, in the evaluation domain, the pattern of "setting up a test scenario → running the test software → reading the test data → drawing conclusions" is repeatedly trained and becomes the model's default "comfort zone." When multiple accounts use the same model, they easily and unconsciously apply this pattern, forming a deep-seated logical homogenization (even if the surface language and topics differ, the core thought process is consistent).

[0042] The specific implementation includes: defining a regular expression for a homogenization logic chain; and pre-defining typical token sequence patterns that may lead to homogenization based on experience and historical data analysis, such as: / test→data→conclusion / g. The regular expression / test→data→conclusion / g means matching a continuous token sequence: a token that roughly represents "test" (such as "test", "actual test", "run benchmark") → a token that roughly represents "data" (such as "result", "display", "measured") → a token that roughly represents "conclusion" (such as "indicates", "visible", "therefore") (g indicates a global match).

[0043] Detection and Masking Operations: When the model begins generating content or each new token is generated, the system scans the existing complete context token sequence. Specific example: Suppose a Brazilian account has generated the following context: "We conducted a heat dissipation test on the iPhone 15 Pro by running the GenshinImpact game. Software data shows...". The scan reveals a test → data sequence. If the next generated token might indicate, conclude, etc. (i.e., potentially completing the test → data → conclusion chain), then when generating that token, it is marked in the attention mask matrix: when predicting this specific potential conclusion token, the original attention weights of the test and data tokens are masked (assigned zero values). The effect is that when the model generates words that might continue the test → data → conclusion pattern, it cannot effectively utilize the attention information of key preceding words in that pattern. This is equivalent to interfering with the model's habitual thought process, thus forcibly breaking the model's dependence on high-frequency homogeneous logical chains, forcing the model to find alternative information associations and expressions. For example, what might have been said as "Test data shows poor heat dissipation, but it can still be used normally" could now be changed to "Although test data shows high temperature, the actual grip in tropical climate is not as hot as expected..." The latter introduces different perspectives such as subjective feelings (grip), regional comparison (tropical climate), and transition logic, which injects differentiation at the level of logical expression.

[0044] From surface language to the diversity of internal structure – dynamic temperature control influences word selection style (surface); attention masking adjusts the attention flow within the model, changing the content organization structure and logical progression (deep); precise and automated – the settings of control parameters (T0, k, regular expression pattern) are directly derived from the character genotype structure, ensuring customized generation strategies for different character types.

[0045] Existing technologies rely on manually writing complex prompts or examples (Few-shot learning) to adjust styles or suppress templates. This solution intervenes through mathematical models and underlying mechanisms (dynamic hyperparameters + attention masking) to guide and constrain model behavior in real time, automatically, and at a deeper level during the generation process, resulting in more stable, controllable, and quantifiable effects.

[0046] S130, the generation environment control file is loaded into the generation model, and differentiated content generation is performed.

[0047] Users ultimately need content text that aligns with their persona, has a clear theme, is rich in content, and is differentiated from one another. Existing content generation methods face two key challenges: information clutter and unclear focus—user-input topics (such as the topic sentence "iPhone 15 Pro heat dissipation performance") may encompass multiple aspects (design, materials, testing, software, comparison), and models may not be able to effectively break them down and capture the key points that best fit the local audience and the account's DNA; and superficial regional adaptation—even with regional identifiers, the model's application of regional characteristics is often shallow (such as replacing city names and currency symbols). True localization requires deeply integrating regional characteristics into the substantive perspective and element selection of the content.

[0048] Therefore, in the implementation, the input includes: user-provided topic text (e.g., "iPhone 15 Pro heat dissipation performance evaluation", which is the starting point for model-generated content), a large model loaded with the generation environment control file (such as GPT-4, Llama), and the regional feature dataset contained in the persona genotype structure (localeFeatures_BR extracted by CLDR).

[0049] An Abstract Meaning Representation (AMR) parser is used to deconstruct the input topic into atomic elements. AMR is a graph-based semantic framework that abstracts and structures the semantics of natural language sentences. It strips away specific words and syntactic structures, extracting the core semantic roles (such as action, agent, patient, time, place, and attribute) and their relationships. Specifically, the AMR parser processes the topic sentence "iPhone 15 Pro heat dissipation performance evaluation". Example of parsing results (simplified AMR graph representation):

[0050]

[0051] The atomic element deconstruction output is the core semantic role nodes extracted from the AMR graph, which become the "atomic elements" for subsequent operations: Device: iPhone 15 Pro; Attribute: Thermal Dissipation - a core evaluation point; Action: Review; (implicit) Agent: Author / Reviewer. The technical value of AMR is that it transforms topic sentences from formalized text into a set of core concepts (Device, Attribute, Action) with clear semantic relationships, laying a clear semantic foundation for subsequent refined processing and differentiated generation, avoiding ambiguity at the word level (e.g., "performance" can refer to multiple aspects).

[0052] The priority sequence of elements is dynamically rearranged based on the regional feature dataset. While the elements deconstructed by AMR are crucial, their importance is not equal and must be adapted to regional characteristics. For example, in hot regions, "high temperature impact" and "outdoor usability" may be dimensions that are closer to user pain points than "chip specification details" when evaluating heat dissipation performance. The weight and presentation order of these elements in content generation need to be dynamically adjusted based on the regional features extracted by CLDR (such as commonTopics:{heat:[...],outdoor:[...]}).

[0053] The specific implementation includes: constructing a regional feature weight vector Wg = (w1, w2, ..., wn), where n represents the number of core elements deconstructed by AMR (at least 4 in this example), and the formula for calculating wi (the regional weight of element i) is wi = Σ_{j=1}^{m}freq(termj) / freq(termi). freq(termi) is the frequency of the literal word or core concept word of element i in the entire corpus (high frequency indicates strong universality, and the weight of regional features should be relatively reduced); freq(termj) is the regional feature (such as the commonTopics.heat array ['calor intenso', '...' in localeFeatures_BR). The frequency of each word in the set {termj, j = 1..m} of the feature words strongly associated with the region (high frequency indicates high regional interest in this topic). The reverse logic is: high freq(termi) -> low wi (weakening general elements); high freq(termj) -> high wi (emphasizing elements associated with the regional interest words). Calculate the element reordering sequence: Snew = argsort(Soriginal × Wg). Soriginal is the original AMR element sequence, such as [Device, Attribute, Action, Agent]. This sorting is usually grammatically driven (subject-verb-object) or semantically driven; × indicates element-wise multiplication (Hadamard product), that is, each original element is assigned its corresponding regional weight wi; argsort(...) calculates the descending index of the weighted new sequence, that is, sorted from largest to smallest according to (Soriginal Wg) values, to obtain the new element priority sequence Snew.

[0054] For a specific example, assuming a Brazilian account: the core element Soriginal = [Device (iPhone 15 Pro), Attribute (Thermal Dissipation), Action (Review), Context (Usage Scenario)] (including the implicit element "Usage Scenario" or an element extended from the topic); the geographic related keywords: commonTopics_heat = ['calor intenso' (freq high), ' quente'(freq high)], commonTopics_outdoor=['praia'(freq high),'parque'(freq medium)], total Σfreq(termj) very high; common frequency: freq("ThermalDissipation") (medium) vs freq("Device") (very high) vs freq("Context") (medium) vs freq("Action") (very high). Calculating Wg (estimated example): For Context: Σfreq(termj) is high (associated with heat and outdoor), freq("Context") is medium -> W_context = high value / medium value = high value; For Attribute: Σfreq(termj) is high (associated with heat), freq("Thermal Dissipation") is medium -> W_attr = high value; For Device: Σfreq(termj) is low (regional feature words are not directly associated with specific device names), freq("Device") is high -> W_device = low value / high value = very low value; For Action: similar to Device, W_action = very low value. Soriginal × Wg ≈ [Device very low value, Attribute high value, Action very low value, Context * high value]. argsort (descending order) => Snew = [Attribute, Context, Device, Action] or [Context, Attribute, Device, Action]. The model guidance effect is that this new priority sequence Snew is integrated into the control information of the generative model (possibly as a cue word modification or internal guidance signal). It tells the model to prioritize the core focus when generating content: first, emphasize the thermal performance (Attribute) itself and its closely related usage scenarios (Context - especially high temperature, outdoor); second, or briefly describe the device description - the introduction of the device (iPhone 15 Pro) can be placed later or briefly mentioned; and simplify the evaluation behavior - the Action (Review) verb does not need to be emphasized.

[0055] Differentiated content generation execution: The model, loaded with a control file (including dynamic temperature and attention mask), begins generating text token-by-token after receiving user prompts on topic and feature priority. The control file ensures: the appearance of CLDR data-driven localized expressions (such as currency symbols and high-temperature vocabulary) in appropriate locations; when discussing "use cases (Context)," the dynamic temperature mechanism may increase T(t), encouraging the generation of more vivid and localized descriptions; when the model instinctively tries to move towards test->data->conclusion, the attention mask intervenes to interrupt; and the content always revolves around the highest priority Attribute (heat dissipation performance) and Context (high-temperature outdoor). Output example (differentiated content snippets that a Brazilian account might generate): "In Rio's 40°C heatwave (regional context), the iPhone 15 Pro's heat dissipation capabilities faced extreme challenges (attribute-focused). Don't worry about it malfunctioning in the Carnival crowds—we tossed it into the sand on the beach to play games... Surprisingly, despite the phone being so hot it felt like it had just been taken out of a sauna (high temperature + subjective experience introduced by dynamic temperature at high points), the frame rate was much more stable than last year's Xiaomi flagship (comparative perspective introduced by attention masking to prevent possible alternative paths after the standard conclusion chain). Is it worth the R$9000 price tag to conquer the Brazilian heatwave? (localized Currency + scene fusion) This is more telling than just looking at lab data (avoiding directly entering the data-conclusion chain)."

[0056] AMR element rearrangement forcibly alters the entry point and narrative structure of generated content. Brazilian accounts approach the topic from the perspective of "high-temperature environments," discussing heat dissipation experiences. German accounts, due to CLDR characteristics and genotype settings, may prioritize "structural design" or "temperature control accuracy" after AMR rearrangement, shifting the entry point to technical analysis (even if the theme is the same: heat dissipation). Substantial localization is achieved—through the Wg formula, regional characteristics deeply shape the content's focus, going beyond mere word substitution. Brazilian content's core concern is "practical use in extreme weather," rather than simple performance metrics. Dynamic temperature and attention masks disrupt the model's conventional expression path and templated logic during generation, resulting in diverse text styles and distinct logics, effectively avoiding core homogenization caused by highly similar input topics (all reviewing iPhone heat dissipation). Through AMR decomposition and regional rearrangement, the substantive perspective, narrative focus, and emotional tone of content produced by accounts from different regions on the same topic show significant differences. Existing technologies may add "high-temperature outdoor applications are a priority" to the prompts. This solution uses semantic parsing (AMR) and graph kernel algorithms (Wg) to structurally, quantitatively, and automatically reconstruct the priority of the core elements of the topic and apply it to the deep generation logic of the model. The overall solution is synergistic and efficient—this step perfectly inherits the results of the first two steps (genotype, control file) and applies them to the core link of the final content generation (element deconstruction and rearrangement), which is the key execution link for the entire solution to achieve differentiated content.

[0057] S140 analyzes the cognitive homogenization risk of generated content through graph isomorphism detection algorithms. Existing content monitoring technologies focus on surface indicators: sensitive word filtering (compliance), interaction volume / completion rate / click rate (user interest), and plagiarism detection (direct copying). These cannot capture core homogenization (e.g., all accounts' iPhone heat dissipation reviews are highly similar in logical structure, viewpoint focus, and argumentation methods). With hundreds of accounts managed by operators, it's impossible to read and compare content one by one, leading to the hidden accumulation of homogenization risks. These risks are only discovered when the platform algorithm lowers the ranking due to content homogenization (e.g., being identified as "machine-generated" or "low-quality template"), by which time it's too late. There is an urgent need for an automated, fundamental homogenization detection method. The implementation method includes: the text of the newly generated content to be published (e.g., the iPhone heat dissipation review generated by the Brazilian account in the previous step), and relevant historical content in the database (including (a) the account's (Brazil's) historical posts; (b) recent content from other 3C review accounts (especially those in the same industry, region / country).

[0058] The process includes: transforming the content into a labeled directed graph G = (V, E). Converting natural language text into a structured semantic graph allows for a deeper capture of the logical framework and conceptual relationships of the content, ignoring superficial lexical variations. This is more effective than simply calculating word vector similarity in capturing the essence. Specific operations involve using Semantic Role Labeling (SRL) or dependency parsing combined with entity extraction. Node V (Vertex) represents the main semantic role entities and core actions / states, such as entities: iPhone 15 Pro, heat dissipation performance, high-temperature environment (40℃ Rio), game (GenshinImpact), perceived temperature, price, previous generation Xiaomi flagship; main actions / states: challenge, use, test, display, compare. Edges (E) represent logical relationships between nodes, usually directional. For example, ARG0 (Agent): iPhone 15 Pro -- Challenge -- Heat dissipation performance; ARG1 (Patient / Target): High-temperature environment -- Impact -- Heat dissipation performance; LOC (Location): Use -- LOC -- Beach sand pile; MNR (Method): Test -- MNR -- Game; TEMP (Time): Display -- TEMP -- Long-term operation; Comparison object: iPhone 15 Pro -- Comparison -- Previous generation Xiaomi flagship; Attribute: Heat dissipation performance -- Attribute -- Perceived temperature. Each node and edge has a semantic label (such as node type, relationship type). Result graph (simplified example):

[0059] "(Challenge (ARG0: iPhone)(ARG1: Heat Dissipation Performance)) -- Condition (COND) --> (High Temperature Environment)"

[0060] (Using (ARG0: Author)(ARG1: iPhone)(LOC: Beach Sandpit)) -- Method (MNR) --> (Play (Game))

[0061] (Display (ARG1: Performance)) -- Comparison (COMP) --> (Previous Xiaomi) (Specific Attribute: Frame Rate)

[0062] (Heat dissipation performance) -- Attribute (ATTR) --> (Feeling temperature)

[0063] (Feeling temperature) -- Description (DESC) -- ("Like being taken out of a sauna")

[0064] Cross-account graph similarity is calculated based on the Weisfeiler-Lehman (WL) graph kernel algorithm. The WL algorithm is an efficient graph isomorphism / similarity approximation algorithm. It iteratively hashes and aggregates the label information of nodes and their neighbors, generating increasingly richer node representations, thereby determining the overall structural similarity of the graph. Core algorithm operations (iterative formula): k is the iteration round (k = 0 is the initial label); v is a node; L^(k)(v) is the label of node v after the kth iteration; N(v) is the set of direct neighbors of node v; | is the concatenation symbol (connecting the current node label with the set of labels of all its neighbors); HASH() is a hash function that maps the concatenated string to a new, higher-level label.

[0065] Calculation process: When k=0, all nodes v obtain their initial semantic labels L^(0)(v) (e.g., "iPhone", "heat dissipation performance", "usage"); when k=1, for each v, collect L^(0)(v) and the L^(0)(u) of all neighbors u, concatenate and hash to generate L^(1)(v), which integrates v's own information and the structure of its first-order neighbors; when k=2, repeat the above process, but the neighbor labels are L^(1)(u), which integrates the structure of second-order neighbors; and so on, until the preset iteration depth K is reached. Feature vector / multiset: After the iteration is completed, record all node labels L^(k)(v) generated in each round k, and count the number of times each label appears in different rounds. For a graph G, the WL feature can be represented as: WL_k(G)={(Label,Frequency initeration k)for k=0..K}. Similarity calculation: sim(G1,G2)=WL_k(G1)∩WL_k(G2) / WL_k(G1)∪WL_k(G2). Formula explanation: Calculate the Jaccard similarity of the feature sets of two graphs (G1 is the new content graph, G2 is the historical content graph) after K rounds of WL iterations; ∩ represents the sum of the counts of features shared by G1 and G2 in all features of K rounds of iterations; ∪ represents the total number of features of G1 and G2 in K rounds of iterations (count after deduplication); WL_k(G) can be understood as the size of the feature multiset; the value of sim(G1,G2) is between 0 and 1, the closer to 1, the more similar the graphs are in structure (semantic skeleton, conceptual relationships).

[0066] Cross-account comparison and risk assessment are automatically calculated by the system: new content vs. the account's history (to avoid self-duplication of account content); new content vs. recent content clusters in the same industry (3C reviews) ((a) cluster comparison by region (all Brazilian reviews); (b) global cluster comparison of the same topic (all iPhone heat dissipation related reviews)). Risk threshold: If the average similarity with any group (or with the most similar individual content) exceeds a preset threshold (e.g., 70%), it is considered to have a high risk of homogenization. The output homogenization risk report includes: detection results (e.g., "Similarity calculation result: 82% (higher than the threshold of 70%)"); similarity point analysis (e.g., "Highly similar core: multiple articles use device → performance attributes → environmental challenges → game performance comparison as the core logical framework, price consideration nodes appear frequently but are similar in expression. Lack of regional culture (e.g., deepening the perspective of carnival / beach crowd interaction) or unique viewpoint branches (e.g., optimization of specific application scenarios)"); adjustment suggestions (e.g., "High risk. Suggestions: (1) Adjust the genotype, reduce the weight of the cost-effectiveness factor or introduce the local service experience factor; (2) Trigger content regeneration, strengthen the description of social scenarios in high temperature environments or specific software optimization experience details; (3) Check whether the homogenization logic chain mask is effective").

[0067] The WL graph kernel algorithm penetrates the surface differences of language, directly comparing the core skeleton of the content's inherent semantic structure and logical relationships, accurately capturing deep homogenization patterns that are difficult for humans to detect; it is preventative rather than reactive monitoring—identifying risks before publication, avoiding post-event losses due to platform penalties; and it provides quantitative indicators to guide operations—offering objective similarity values ​​so that operators understand the degree of risk and clarify optimization directions. Its advantages over existing technologies are: surpassing word vectors and NLP surface similarity—word vectors (BERT embeddings) calculate cosine similarity, focusing more on the degree of semantic matching, and are easily deceived by superficial word substitutions; the WL graph kernel focuses on the similarity of structural relationships, which is the essential method for solving the core homogenization problem; and it automates batch processing—achieving automated cross-comparison between massive amounts of content, solving problems that cannot be covered manually.

[0068] S150, when the graph similarity exceeds the preset threshold, the persona genotype is updated. Existing technologies lack a closed-loop feedback mechanism based on the inherent risks of the core content. In the implementation method (detailed according to the technical solution), the input is the high-risk judgment and homogenization risk report in step (4). The trigger condition is sim(G_new,G_ref)>Threshold (preset threshold). Updating the persona genotype structure (core operation) includes: removing nodes with an overuse rate higher than the threshold value in the industry-specific entity set. The industry-specific entity set defines the core focus of the account. Some nodes (such as cost-effectiveness) are detected as overused in the content, becoming one of the "culprits" leading to homogenization (for example, all 3C reviews repeatedly emphasize cost-effectiveness, ignoring other unique points). Specific operations: Calculate overuse rate - Calculate the frequency of the entity (cost-effectiveness) appearing in the recent historical content of the account, compare it with the average frequency of other entities in the industry-specific entity set, and calculate a "relative overuse rate", or simply set an absolute frequency threshold; Removal operation - If the overuse rate of the entity exceeds the set threshold, it will be temporarily removed from the industry-specific entity set (core entity set) or significantly demoted (e.g., Brazilian account core entity set: remove [cost-effectiveness] or mark it as deprecated).

[0069] Entities are randomly selected from a cross-domain knowledge base and injected into the entity set. Simply removing overused nodes may result in a narrow content scope. Injecting new entities can force the model to broaden its perspective, generating unexpected and differentiated content connections. Specific operations: The cross-domain knowledge base is a broader knowledge base that may contain fragments of knowledge graphs related to technology, entertainment, sports, and lifestyle; random selection is not completely random but based on certain rules (e.g., selecting entities with a weak correlation to the current industry, avoiding completely unrelated entities; combining certain preference coefficients existing in the genotype; setting an upper limit on the number of injected entities); Injection example: For a Brazilian account, randomly select entities to inject: social impact (e.g., the embarrassment of the camera becoming unusable due to overheating while taking photos on social media), music festival battery life (indirectly related to heat dissipation: temperature affecting the battery causing shutdown and missing the performance), localized applications / games (e.g., the performance of a hardware-intensive local game popular in Brazil, *LoL: Wild Rift*, under high temperatures). Example of the updated entity set (Brazil): Core entity set = [heat dissipation performance, battery life, outdoor use, high temperature environment, local warranty, gaming performance, social impact, music festival battery life, localized application / game] (representing newly injected cross-domain entities).

[0070] It also adjusts genotype parameters—dynamically updates the weights of corresponding parameters in the genotype (e.g., local service experience coefficient += 0.2) based on risk report recommendations (such as enhancing local service experience); and regenerates content—generating new, differentiated content based on the updated genotype structure.

[0071] The newly generated content undergoes homogenization testing again, while the actual platform performance metrics of the account are monitored (such as recommendation volume, exposure volume, and interaction rate). If the risk of the new content decreases and the performance metrics improve or remain the same, the update is effective; otherwise, it may trigger further genotype adjustments or parameter optimizations. Reinforcement learning optimization (extension) – The system can design a reward signal (such as the percentage increase in recommendation volume minus the percentage decrease in homogenization risk) and use reinforcement learning algorithms (such as PPO) to optimize the update strategy of genotype parameters (such as T0, k, entity selection weights) to achieve long-term adaptability.

[0072] Starting with defining the genotype of core personality attributes, removing overused entities that cause problems, and injecting new elements, the direction of content generation is fundamentally adjusted; continuous evolution—through closed-loop feedback (detection -> update -> generation -> re-detection), the account's personality and its content can dynamically adapt to environmental changes (platform algorithms, competitor strategies, shifts in user interests), maintaining differentiation and competitiveness in the long term; automated maintenance—greatly reducing the workload of manually adjusting genotypes and supporting the management of massive numbers of accounts.

[0073] Existing technologies define character profiles statically or adjust them solely based on interactive data (the latter only reflects the result and cannot prevent core homogenization). This solution performs preventative and precise genotyping based on the core homogenization detection results of the content semantic structure; it introduces cross-domain characteristics—randomly injecting cross-domain entities, which is a creative coercive strategy that is difficult to simulate by human design or rule engines, effectively expanding the boundaries of content.

[0074] The overall processing flow (a complete closed loop for a single new piece of content) includes:

[0075] Initialize / request: The user or system scheduler requests the generation of a new article for a specified account (e.g., ID: BR_3C_001) with the subject: "iPhone 16 heat dissipation performance review".

[0076] Step (1) - Genotype Construction: The system reads the preset target region identifier BR and industry classification identifier Consumer_Electronics_Review based on the account ID; Execution: Call CLDB to extract localeFeatures_BR; Execute knowledge graph pruning SPARQL query (using the optimized PageRank threshold) to obtain or update the core entity set; Output: Current person genotype structure HGT_BR_3C_001 (containing regional features, core entity set, and possible parameters such as T0,k).

[0077] Step (2) - Control file construction: Based on HGT_BR_3C_001: Determine the dynamic temperature function parameters (T0 = 1.0, k = 0.3, Tc = 60); Generate attention mask rules (e.g., still using / test→data→conclusion / g mode); Output: Generate environmental control file CEF_BR_3C_001.

[0078] Step (3) - Differentiated Content Generation: Load the user topic "iPhone 16 Heat Dissipation Performance Evaluation"; Execute: AMR parser processes the topic -> obtains the atomic feature set: [Device(iPhone 16), Attribute(ThermalDissipation), Action(Review), Context(Implicit)]; Load the localeFeatures_BR -> calculate the local weight vector Wg -> feature rearrangement: [Attribute,Context,Device,Action] (assuming high regional correlation); Load the model + CEF_BR_3C_001 + feature priority; Output: Generate preliminary content draft Content_Draft_001.

[0079] Step (4) - Homogenization Risk Detection: Execution: Convert Content_Draft_001 into a semantic graph G_new; Query and obtain the relevant historical content graph set RefGraphs: the historical graph of this account (such as the past 5 pieces of content), the recent content graphs of other Brazilian 3C accounts (such as the top 10 likes), and the recent global iPhone heat dissipation related content graphs (such as the top 20 popularity); For each G_ref in RefGraphs: sim_k = WLSimilarity(G_new, G_ref) (K = 3); Calculate the average similarity (or maximum value): AvgSim_Brazil = 0.75, AvgSim_Global = 0.68; Determination: AvgSim_Brazil(0.75) > Threshold(0.70) -> High risk!; Output: High risk report (indicating high similarity to the logical framework of Brazilian counterparts, similar game performance, and similar price node patterns).

[0080] Step (5) - Genotype Update and Regeneration: Trigger: Due to high risk judgment; Perform genotype update: Remove overused entities - Analyze historical content and find that the cost-effectiveness entity appears abnormally frequently and has similar expressions in recent Brazilian content (overuse rate exceeds the limit), remove operation: HGT_BR_3C_001.core entity set.Remove('cost-effectiveness') (or reduce weight); Inject new entities - Randomly select entities with weak correlation to social, outdoor, and local culture from cross-domain knowledge base (such as live broadcast effect, dust protection, regional festival needs), inject operation: HGT_BR_3C_001.core entity set.Add(['live broadcast effect', 'dust protection', 'regional festival needs']); Optional) Adjust parameters - According to the report recommendation (strengthen local scene), increase the local scene integration parameter in the genotype; Regenerate content: Repeat steps (1)-(3)->(4): Step (1) uses the updated genotype (including new entities, excluding cost-effectiveness); When generating in step (3), after AMR deconstruction, the regional rearrangement may be affected by entity changes and parameter adjustments. The priority of Context (use scenario) and new entities (live broadcast effect / dust protection) will be further increased, and new content will be generated under the guidance of the control file; Step (4) detect the new draft Content_Draft_002, assuming that AvgSim_Brazil = 0.58 < 0.70 -> low risk. Output and release: Content_Draft_002 is marked as low risk and enters the release queue or is released after review. Monitoring and long-term optimization: Monitor the platform metrics (recommendation volume, interaction) of the BR_3C_001 account after release. If the metrics decrease or homogenization accumulates again, it will be used as a signal input to the reinforcement learning module to optimize the genotype update strategy in step (5) (such as adjusting the overuse threshold, injecting entity selection rules, and parameter adjustment range).

[0081] Application Example (3C Review Account Group Management):

[0082] A Brazilian account generated "High Temperature Carnival: A Real-Life Account of iPhone's Heat Dissipation Survival on a Rio Beach" (structure: scene narrative + climate analogy); a German account generated "Thermodynamic Anatomy: Precision Errors in the iPhone's Heat Dissipation Structure" (structure: experimental deduction + negative feedback analysis); and a Japanese account generated "The Philosophy of Cooling and Silent Operation: How to Achieve Heat Dissipation in Tiny Spaces" (structure: dialectical contradiction + cultural metaphor). A key indicator of technological gap: While sharing the same theme, the three accounts share only a 12% similarity in logical structure (compared to a traditional approach of >65%), indicating a complete divergence in cognitive patterns. Specifically, the Brazilian account (BR_3C_001): Initial genotype: Region BR, Industry Consumer_Electronics_Review; Parameter settings: Cultural integration coefficient = High, Innovation index = Medium; CLDR focuses on high temperature and outdoor activities; Theme generation: "iPhone 15 Pro heat dissipation performance review"; Differentiated generation results (after processing): "Testing the iPhone 15 Pro under the heat wave of the carnival crowd (high cultural integration) (Context first), the heat dissipation performance is your samba rhythm stabilizer (localized metaphor, innovation index driven)! Beach test: The body is scorching hot? Indeed! But I'm more afraid that it will freeze like last year's Android phones and become like a PPT (avoiding standard data conclusion chain, injecting subjective + comparative perspective). Will it overheat and shut down during the live broadcast? (Newly injected physical live broadcast effect) We will help you burn it for 48 hours (...)"; Core characteristics: Localized scene entertainment evaluation (focusing on the interesting description of the actual use scenario at high temperature, while taking into account live broadcast concerns), which is significantly different from the core of the German / Japanese account. German account (DE_3C_001): Initial genotype: Region DE, Industry Consumer_Electronics_Review, Parameter settings: Innovation Index = High, Controversy = Low, CLDR implies engineering and precision; Related topic generated results: "iPhone 15 Pro Cooling System Engineering Analysis (In-depth Technology Guide): Microcirculation of the Heat Sprayer Under a Microscope (Innovation High Point T(t) Application, Introducing a Professional Perspective). We quantified the heat flow path efficiency under different load gradients (rigorous method), and the precisely designed redundant heat sink is crucial for long-term performance degradation (focusing on technical details and design), and the temperature control under non-extreme gaming scenarios is satisfactory (robust conclusion)."; Core features: In-depth analysis of engineering technical details (emphasizing the principle and quantitative analysis of the heat dissipation structure itself). Japanese account (JP_3C_001): Initial genotype: Region JP, Industry Consumer_Electronics_Review, Parameter settings: Topic preference = Energy saving, CLDR association fine and optimized; Same topic generated results: "Energy saving optimization of iPhone 15 Pro heat dissipation strategy (topic preference guidance): Analysis of how iOS temperature control logic balances heat dissipation and power consumption (local fine preference)".Real-world testing shows a significant reduction in heat dissipation energy consumption under light use (from an energy-saving perspective), and a daily commute temperature control guide in battery protection mode is provided (practical advice-oriented). Core feature: balancing heat dissipation and energy consumption optimization for user experience (emphasizing energy-saving strategies and practical guidelines in everyday scenarios). Homogenization test results: All three articles involve heat dissipation, but their core AMR graph structures (node ​​centroids, core relationship paths) differ: Brazil (environmental challenges -> actual experience -> live streaming concerns), Germany (engineering structure -> principle analysis -> quantitative indicators), and Japan (energy consumption optimization -> balancing strategies -> practical advice). The cross-account similarity calculated by the WL algorithm is far below the threshold (e.g., <50%), the system determines it to be low-risk, and no genotype update is required.

[0083] By generating a persona genotype structure, based on target regional and industry classification identifiers, it accurately captures regional cultural context (e.g., Brazilian users' preference for outdoor high-temperature scenarios) and industry knowledge boundaries (e.g., the potential for differentiation in 3C testing where "energy consumption optimization" takes precedence over "heat dissipation"), thus eliminating redundant information and initial persona ambiguity caused by traditional static prompts. This lays a fundamental foundation for differentiated content generation. Based on the persona genotype structure, a generation environment control file is constructed. By systematically adjusting key parameters of the generation model (e.g., creativity weights and logical structure constraints), it avoids the style rigidity and content template-based approach resulting from existing technologies relying on one-time prompts (e.g., all technical evaluations using a manual-style structure), dynamically guiding the content generation process towards deep localization. After loading the generation environment control file into the pre-trained generation model, differentiated content is generated according to the task theme. This involves deconstructing mixed information on the theme (e.g., focusing on material design rather than generalities in iPhone heat dissipation testing) and incorporating the essence of regional characteristics. Perspectives (such as integrating the high-temperature environment of Brazil into the performance demonstration logic rather than merely replacing place names) are used to achieve dual optimization in content clarity and depth of regional adaptation. A graph isomorphism detection algorithm is used to analyze the cognitive homogenization risk of generated content, using the similarity of the content's internal logical structure graph as an indicator (e.g., comparing different evaluation methods such as the "test → data → conclusion" chain). This automatically identifies core homogenization that surface indicators (such as plagiarism rate) cannot capture (e.g., all accounts have highly similar viewpoints and arguments), thus revealing hidden risks at their essence. When the graph similarity exceeds a preset threshold, a persona genotype structure update is triggered, correcting genotype deviations from the source (e.g., strengthening industry knowledge boundaries to highlight differentiation potential). This forms a closed-loop feedback mechanism, preventing the accumulation of homogenization patterns in subsequent content, thereby systematically improving the uniqueness and localization quality of the content, avoiding platform algorithm demotion due to homogenization, and ultimately achieving efficient generation of highly differentiated and credible virtual persona content, enhancing user experience and content dissemination.

[0084] Example 2:

[0085] Example 2 focuses on a case study of a female farm worker in Texas, USA. This is a complete virtual persona generation process designed for social e-commerce platforms such as TikTok or Instagram. The persona is defined as a young white female farm worker living in Texas, USA. Her daily work includes feeding horses, doing farm work, and using Chinese-made agricultural machinery. She has a humorous personality and a distinct Southern American accent. The goal is to generate high-quality localized content such as nicknames, bios, and posts to improve account recommendation and user engagement. Example outputs include the nickname "Lauren Elizabeth", the bio "let me show you my riding skills", the tag "horse girl, farm", the pronoun "female", and post content such as showing humorous spelling errors like "wandering".

[0086] In this application case, the persona is defined as follows: the region is US-TX (Texas, USA), the industry category is Rural_Lifestyle (specifically, agricultural promotion), and the task theme is "Generate a humorous post showcasing daily farm work." The output content must conform to Texas culture, such as reflecting Southern accents, spelling errors, and localized metaphors.

[0087] Step 1: In response to the input task topic, target region identifier, and industry category identifier, generate a persona genotype structure based on the target region identifier and the industry category identifier.

[0088] This step is the starting point of the process. The system receives input parameters (task topic, region identifier US-TX, industry classification identifier Rural_Lifestyle) and then generates a persona genotype structure (HGT). The persona genotype structure is a data structure (such as a Python dictionary or JSON object) that stores key features: a region feature dataset and an industry-specific entity set, providing a basic template for subsequent steps.

[0089] First, the system uses CLDR to extract regional feature datasets: The system uses the Common Locale Data Repository (CLDR) API, inputting the US-TX identifier, to extract Texas-specific feature data. Specifically, CLDR calls the Unicode database to retrieve fields for the US-TX region, including language habits (Southern English dialects), lifestyle patterns (such as rodeo festivals and high-temperature weather data), and cultural taboos (such as avoiding urban bias).

[0090] Example output: The dataset contains a list of high-frequency words (e.g., "y'all" instead of "you all"), consumer behavior data (agricultural machinery purchase preferences), and dynamic buzzword weights (e.g., the timeliness weight of the "horse girl" tag on TikTok is 30%). Then, the industry knowledge graph is retrieved based on industry category identifiers: the system retrieves relevant subgraphs for the Rural_Lifestyle category from a pre-built industry knowledge graph (built based on Wikidata or a proprietary database). Specific execution: A SPARQL query is executed, matching nodes for "farm work, livestock feeding, agricultural machinery." The initial graph contains 30 entity nodes (e.g., "horse care," "tractoruse," "crop cycles").

[0091] Finally, pruning is performed to create an industry-specific entity set: the importance of nodes is calculated using an optimized PageRank algorithm, with a threshold set to 0.2 (nodes with importance below this value are pruned). For example, the "crop cycles" node, with its low importance, is removed (only highly relevant nodes are retained). Core entities include "livestock feeding," "equipment maintenance," and "outdoor activities." After pruning, an industry-specific entity set (list structure) is formed and stored in a persona genotype structure.

[0092] The example output is: HGT_US_TX_001 = {Local Feature Dataset: [localeData_US_TX], Industry-Specific Entity Set: ["horse care", "equipmentreview", "humor style"], Dynamic Parameter: T0 = 1.2 (Initial Temperature Base Value)}.

[0093] The character genotype structure is dynamically generated based on regional and industry identifiers, ensuring that content is infused with localized features from the source. The regional feature dataset provides cultural details (such as the southern accent of Texas), while the industry-specific entity set focuses on core industry concepts, avoiding interference from unrelated entities. Pruning operations simplify the graph and improve subsequent processing efficiency.

[0094] The output of this step (HGT) serves as the input for the next step, used to build the control file. Existing technologies suffer from the problem that general-purpose models, when generating content, lack structure and localized features (e.g., using CLDR), leading to ambiguous output. For example, when generating nicknames for Texas women, GPT-4 might output the generic name "Texas Girl" instead of the culturally relevant "Lauren Elizabeth"; and when industry graphs are not pruned, the model generalizes content, such as adding irrelevant entities to reduce local relevance.

[0095] In this case study, the nickname "Lauren Elizabeth" is taken from a common Texas naming pattern, and the personal profile "let me show you my riding skills" directly references regional data, reflecting Texas's horseback riding culture and thus enhancing user resonance. Compared to traditional solutions (such as web scraping accounts for content creators), this solution is superior because it dynamically generates a persona genotype, rather than a static prompt; existing AI can only generate 50% cultural relevance, while this solution achieves over 90% localization accuracy through CLDR and pruning.

[0096] Step 2: Based on the human character genotype structure, construct and generate the environment control file.

[0097] This step uses the persona genotype structure (HGT_US_TX_001) generated in the previous step to construct the generation environment control file (CEF). This file contains a dynamic sampling temperature function and an attention mask matrix, which are used to regulate the large model generation process and ensure output diversity.

[0098] First, determine the information entropy of the industry-specific entity set: calculate the information entropy of the industry-specific entity set. Specifically, statistically analyze the category distribution of the entity set (e.g., "horse care" entities account for 40%, "equipment" for 30%, and "humor" for 30%). The entropy formula is: H = -Σ(p_i logp_i), where p_i is the category percentage. The calculated entropy value in the case study is H = 1.5 (moderate dispersion), indicating strong industry knowledge diversity (the threshold is set at entropy > 1.0 to indicate high diversity).

[0099] Next, a dynamic sampling temperature function is configured that fluctuates with the generation time cycle: the function parameters are adjusted based on the entropy value. The formula for the dynamic sampling temperature function is: T(t) = T0 + k sin(2πt / Tc), where T0 is the base parameter and k is the fluctuation amplitude coefficient. The parameters are set according to the entropy value: k = 0.4 (directly proportional to the entropy value H = 1.5), T0 = 0.8 (inversely proportional to the entropy value), and Tc = 50 (generation cycle 60 seconds). This function regulates randomness in the generation model: increasing content innovation at high temperatures and reducing risk at low temperatures. Then, high-frequency logical chain patterns are statistically analyzed and an attention mask matrix is ​​constructed: entity sequence chains (logical relationships between entities) are analyzed from industry-specific entity sets.

[0100] Specific execution: Frequency statistics of sequence chains are performed (e.g., "feed animals->use tractor" appears frequently). Chains exceeding a threshold of 0.4 are marked as high-frequency patterns (e.g., "horse care->ridingskills" has a frequency of 0.6). An attention mask matrix is ​​constructed: a two-dimensional matrix where the weights of sequence chains matching high-frequency patterns are set to 0, suppressing the output of these word sequences in the large model's attention mechanism. Example output: In the mask matrix, the weight of the pattern "care->skills" is set to zero to avoid generating duplicate content. Finally, a generation environment control file is constructed: integrating the dynamic temperature function and the mask matrix to form the control file CEF_US_TX_001, in JSON format, for loading into the large model.

[0101] The generation environment control file regulates the large model through dynamic functions and mask matrices. The dynamic sampling temperature function is based on information entropy: increasing the k value enhances innovation when entropy is high (high diversity), and decreasing k when entropy is low to avoid risk; the mask matrix suppresses high-frequency logic chains, breaking the homogenization pattern. The principle ensures a balance between local innovation and global difference in content. Logical continuity is reflected in the fact that the CEF file output in this step is the input for the next content generation step, used to control the model output. Existing technical problems: The technical problem is that existing large models (such as GPT-4) lack a temperature function fluctuation mechanism, resulting in high content similarity (e.g., all agricultural accounts post "Using farm tools report"). High-frequency logic chains are not suppressed, leading to cross-account kernel similarity >70%. For example, without masking, the large model repeatedly outputs the "feeding horses -> testing machinery" sequence, which is easily penalized by the platform's algorithm.

[0102] Dynamic functions increase randomness and enhance the proportion of innovative content; masking matrices reduce the usage rate of high-frequency patterns, implying a 50% reduction in the risk of content homogenization. In the case study, the personal profile "let me show you my riding skills" avoids excessive use of the high-frequency word "riding" under the masking, replacing it with a humorous style to improve localization.

[0103] Step 3: Load the generated environment control file into the pre-trained generative model and generate differentiated content based on the task theme.

[0104] This step uses CEF files to control the large model and generate differentiated content. For the task theme "Generate a humorous post showcasing daily farm work," the following sub-steps are performed: First, obtain the raw corpus based on the task theme: The system parses the task theme into raw text, such as "Humorous Post: Sharing Daily Farm Horse Riding Experiences," as input corpus. Second, use an Abstract Semantic Representation (AMR) parser to decompose semantic elements: The AMR parser (based on an open-source framework such as CAMR) decomposes the raw corpus into a set of semantic elements.

[0105] Specific execution: The input text is parsed into nodes and edges (e.g., action node: Show, attribute node: Humorous, entity node: Horse, context node: Farmwork). Example output: The semantic element set contains ["Action:Show","Attribute:Humorous","Entity:Horse","Context:Farmwork"]. Then, the semantic elements are weighted according to the frequency of use of specific regional terms: the weighting system uses the formula W_new = (α × timeliness) + (β × regional coverage) - (γ × conflict index).

[0106] The parameters are based on a regional dataset: α = 0.5 (timeliness, e.g., "horse girl" is a new buzzword), β = 0.7 (regional coverage, e.g., "rodeo" is highly relevant in Texas), γ = 0.3 (conflict index, avoiding negative words). Element weights are calculated: for example, "Entity:Horse" has a weight of 0.8, and "Context:Farmwork" has a weight of 0.7. After weighting, the elements are sorted by weight and output as: ["Attribute:Humorous","Entity:Horse","Action:Show","Context:Farmwork"]. Next, filtering and substitution are performed: the original input corpus is filtered through a religiously sensitive word database and a regional folklore word database.

[0107] Specific execution: The dictionary is matched (e.g., to avoid conflicts with "citylife"), and matched words are mapped to alternative words (e.g., "urban" replaces "city"). Example output: A filtered set of semantic elements, retaining words like "wandering" (intentionally misspelled to reflect colloquialism). Finally, the input to the generator model outputs differentiated content: The sorted sequence of semantic elements is loaded into a pre-trained large model (e.g., a localized model fine-tuned based on GPT-4), and the CEF file controls the process: a dynamic temperature function increases randomness, and a mask matrix suppresses high-frequency sequences. Generated content output: The text post "Still wandering why I haventheard from you yet?" Honestly, it gets a little lonely sometimes—whether I'm lounging in a book at the library or curled up in bed…, matching a southern accent and humor. Simultaneously generating multimodal content: an image showing a horseback riding scene (generated based on text feature alignment).

[0108] AMR parsing ensures accurate semantic decomposition; weighting and sorting are based on regional characteristics to enhance local relevance; filtering and substitution avoid cultural conflicts. Dynamic functions in control files inject innovation (such as humorous misspellings of "wandering"), and masks prevent identical sequences. The principle is to prioritize high-weight elements through sorting logic to ensure consistent character portrayals.

[0109] Existing large-scale models do not utilize AMR (Adaptive Character Recognition) and weighting, resulting in a high rate of cultural symbol misuse (e.g., GPT-4 outputs formal spellings, lacking localized errors) and short-lived character consistency (<3 rounds of dialogue); without filtering, the violation rate is 12.1%. Through weighting and substitution, the derived localization acceptance rate reaches 89.7%, and the violation rate drops to 0.3%. In the case study, the "wandering" error in the post reflects the online habits of Texas youth, increasing user resonance. Traditional AI solutions only mimic the output, achieving a character creation efficiency of 200 characters / hour; this solution, through file control and weighting, achieves an efficiency of 2000 characters / hour, a 900% improvement.

[0110] Step 4: Analyze the risk of cognitive homogenization of differentiated content using graph isomorphism detection algorithms.

[0111] This step checks whether the generated content has a homogeneous core, avoiding the problem of superficial compliance but similar structure.

[0112] First, represent the differentiated content as a labeled directed graph: use a semantic role annotation tool (such as PropS annotator) to transform the content text into a labeled directed graph. Nodes in the graph represent semantic roles (e.g., entity "horse", action "show"), and edges represent logical relationships (e.g., "horse--association -> ridingskills"). Example: Convert post content to graph G_new = {nodes: ["Action:Show", "Entity:Horse", "Attribute:Humorous"], edges: ["association", "causation"]}.

[0113] Then, similarity is calculated based on graph isomorphism detection algorithms: the Weisfeiler-Lehman (WL) algorithm (preset K = 3 levels) is used to compare the similarity between G_new and the reference graph set. Reference graph sources include: historical content graphs of this account (e.g., the past 3 posts), graphs of other Texas farm accounts (Top 10 local hot posts), and global agricultural equipment content graphs (Top 20 most popular posts). The similarity formula is: sim = the ratio of matching nodes to edges. Case calculation: similarity with the Brazilian account graph is 0.65, and similarity with the global content graph is 0.70 (preset threshold 0.70). Finally, the risk of cognitive homogenization is determined: if the average similarity > the threshold (0.70 ≥ 0.70 in this case), it is considered high risk; otherwise, it is low risk.

[0114] Graph structures capture core logic (rather than surface text), and the WL algorithm efficiently detects isomorphism (similar structures). High similarity indicates core content similarity, making it susceptible to demotion by platform algorithms. Existing technologies monitor surface metrics (such as keywords), ignoring structural homogenization, leading to a silent decline in content weight. For example, a post may be compliant on its own, but its core content may be similar to global content.

[0115] Graph detection improves the early detection rate of homogenization risks by 80%, preventing account throttling. In the case study, a risk alert was triggered when the similarity was 0.70.

[0116] Step 5: If the graph similarity of the cognitive homogenization risk exceeds the preset threshold, then update the persona genotype structure.

[0117] This step dynamically updates the genotype when high-risk cases are detected (e.g., case similarity of 0.70), achieving closed-loop optimization.

[0118] First, calculate the overuse rate metric: Based on the risk results, analyze the industry-specific entity set in the persona genotype structure. Calculate the historical usage frequency of each entity node across multiple accounts (e.g., "humor style" appears 70% of the time across 100 accounts). Calculate the overuse rate formula: frequency value. Example: Overuse rate metrics include ["horse": 0.85 (high)] and ["tractor": 0.75].

[0119] Then, a threshold value is determined based on information entropy and risk: information entropy H = 1.5 as a reference, risk similarity 0.70, and the threshold value formula is: threshold = risk value × (1 - H / H_max) (set H_max = 2.0). The calculated threshold is 0.65, and entities with an overuse rate exceeding this value need to be removed.

[0120] Next, remove overused entities and inject new entities: Remove nodes with excessive overuse rates from the industry-specific entity set (e.g., remove "horse" with an overuse rate of 0.85), and randomly select new entities from the pre-built cross-domain knowledge base (e.g., "weatherresilience" and "festival event"). Injection logic: Use the API to call the knowledge base and randomly extract weakly related entities. The updated genotype is: HGT_US_TX_001_v2 = {regional features unchanged, industry-specific entity set: ["equipmentreview", "humor style", "weather resilience"]}.

[0121] Finally, regenerate content: loop back to steps 2-4, generating content using the new genotype. Example output: the new post emphasizes "weatherimpact on riding," with kernel differentiation (similarity reduced to 0.55).

[0122] The update mechanism is based on overuse rate and information entropy. It removes high-frequency entities and adds innovative elements to break the cycle of homogenization, so as to achieve long-term differentiation through dynamic evolution.

[0123] The weighting system is implemented in step 3. The W_new formula is used to weight semantic elements. For example, the conflict index is calculated based on the word association degree (in this case, "farmwork" has low conflict).

[0124] Sensitive word filtering and substitution are performed in step 3, filtering and mapping sensitive words such as "city bias" to ensure content security.

[0125] Information entropy calculation is performed in step 2. When the entropy value is high (such as case H=1.5), the entity distribution is discrete, and the derivation is highly diverse.

[0126] The dynamic sampling temperature function is configured in step 2, based on entropy adjustment k and T0 (in this case, k = 0.4, which is proportional to entropy).

[0127] Overall processing flow and application examples (based on a case study):

[0128] The overall processing flow (a complete closed loop for generating a single piece of content) is as follows for the Texas Farm Woman account (ID: US_TX_Farm_001), with the task theme: "Generate humorous posts showcasing daily horseback riding." The initialization request is initiated by the user or scheduler inputting the task theme, the geographic identifier US-TX, and the industry category Rural_Lifestyle; the system assigns an account ID, initiating the closed loop. The genotype construction step involves calling CLDR to extract the US-TX dataset (including language habits and festival culture); querying and pruning the Rural_Lifestyle knowledge graph to form an initial HGT structure (the core entities include "horse care" and "humor"), outputting HGT_US_TX_001. The control file construction step calculates the entity set entropy value H = 1.5; configures a dynamic temperature function such as T(t) = 0.8 + 0.4sin(t / 50); identifies the high-frequency chain "care->skills" to construct a mask matrix, and outputs a CEF file. The differentiated content generation step uses AMR to parse the topic into semantic elements; weighted ranking is applied ("humorous" has the highest weight); conflicting words are filtered; the large model outputs the initial post (containing "wandering" errors) under CEF control, and the content draft is completed. The homogenization risk detection step converts the content into a semantic graph G_new; the WL algorithm compares the historical graphs and finds an average similarity of 0.70 > the threshold of 0.70, indicating high risk. The genotype update and regeneration step statistically analyzes the frequency of the entity "horse" to be 85% > the threshold of 65%; "horse" is removed, and a new entity "weather resilience" is injected; the HGT is updated to V2; after iterative regeneration, the new draft's similarity is 0.55 < the threshold, indicating low risk. The output and publishing stage publishes the final post to the social media platform, with attached images (generated with cross-modal alignment). The monitoring and optimization step tracks platform metrics (such as like rate); signals are input into the reinforcement learning model (setting Reward = 0.7 like rate + 0.3 * conversion rate) to optimize the genotype update strategy (such as adjusting the threshold).

[0129] The application example recreates a Texas account (US_TX_Farm_001): Initial genotype is set to US-TX, industry to Rural_Lifestyle, with parameters including Innovation Index = High and Cultural Integration Coefficient = Strong; CLDR focuses on Southern accents and farm work scenarios. The task theme is generated as "Showcasing daily riding skills." The final differentiated output includes: nickname "Lauren Elizabeth" (derived from common Texas names); bio "let me show you my riding skills" (masking optimization avoids generic words); and post "Still wandering why I haven't heard from you yet?" Honestly, it gets a little lonely sometimes..." (reflecting misspellings and humor, with the core integrating usage scenarios and weather influences); tags #horse girl, #farm, #weather resilience (new entity injection). Core features manifest as localized humor scene evaluation (emphasizing visual entertainment and colloquialisms), distinguishing it from other regions. Data comparison with other accounts (avoiding homogenization): for example, German agricultural accounts output engineering diagrams (in-depth technical analysis of mechanical structures); Japanese lifestyle accounts output practical guides (energy-saving optimization suggestions). Homogenization detection results show cross-account diagram similarity <50% (low risk), ensuring the platform's recommendation weight does not decrease.

[0130] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for generating a digital human virtual persona, characterized in that, include: In response to the input task topic, target region identifier, and industry classification identifier, a persona genotype structure is generated based on the target region identifier and the industry classification identifier; Based on the aforementioned human genotype structure, an environment control file is constructed and generated. The generated environment control file is loaded into the pre-trained generative model to generate differentiated content based on the task theme; The risk of cognitive homogenization of the differentiated content is analyzed using a graph isomorphism detection algorithm. If the graph similarity of the cognitive homogenization risk exceeds a preset threshold, then the persona genotype structure is updated; The process, in response to the input task topic, target region identifier, and industry classification identifier, generates a persona genotype structure based on the target region identifier and the industry classification identifier, including: In response to the input task topic, target region identifier, and industry classification identifier, CLDR is invoked to extract the regional feature dataset corresponding to the target region identifier; Based on the industry classification identifier, retrieve the target knowledge graph corresponding to the industry classification identifier from the pre-constructed industry knowledge graph; Based on the importance of nodes in the target knowledge graph, a pruning operation is performed on the target knowledge graph to delete non-core entity nodes and retain core entity nodes, forming an industry-specific entity set. Store the regional feature dataset and the industry-specific entity set into the persona genotype structure; The process of constructing and generating an environment control file based on the aforementioned human genotype structure includes: Determine the information entropy of the industry-specific entity set; The dynamic sampling temperature function fluctuates with the generation time period based on the information entropy configuration; Statistically analyze the frequency of occurrence of sequence chains consisting of multiple entities within the industry-specific entity set; Entity sequence chains that exceed the frequency threshold are identified as high-frequency logic chain patterns to be suppressed; An attention mask matrix is ​​constructed based on the high-frequency logical chain pattern, and the attention mask matrix is ​​used to suppress word sequences that match the high-frequency logical chain pattern. The generated environment control file is constructed based on the dynamic sampling temperature function and the attention mask matrix; If the graph similarity of the cognitive homogenization risk exceeds a preset threshold, the persona genotype structure is updated, including: If the similarity corresponding to the cognitive homogenization risk exceeds a preset threshold, the historical usage frequency of each entity node in the industry-specific entity set stored in the persona genotype structure within the cross-account range is statistically analyzed to obtain the overuse rate index. The threshold value is determined based on the information entropy and the risk of cognitive homogenization. Remove entity nodes in the industry-specific entity set whose overuse rate index exceeds the threshold value; New entity nodes are selected from the pre-built cross-domain knowledge base and injected into the industry-specific entity set to complete the update of the persona genotype structure.

2. The method as described in claim 1, characterized in that, The step of loading the generation environment control file into the pre-trained generation model and generating differentiated content based on the task theme includes: Obtain the original corpus based on the task theme; The original corpus is decomposed into semantic elements using an abstract semantic representation parser to obtain a set of semantic elements. Weights are assigned to semantic elements in the semantic element set that are associated with the specific regional term based on the frequency of use of the term. The semantic elements in the semantic element set are arranged according to the assigned weight values ​​to obtain a semantic element sequence; The semantic element sequence is input into the generation model, and the randomness of the generation process of the generation model is adjusted by the dynamic sampling temperature function to output differentiated content.

3. The method as described in claim 2, characterized in that, The analysis of the cognitive homogenization risk of differentiated content using the graph isomorphism detection algorithm includes: The differentiated content is represented as a labeled directed graph consisting of semantic role nodes and logical relationship edges between nodes; Based on a preset graph isomorphism detection algorithm, the similarity between the labeled directed graph corresponding to the differentiated content and the labeled directed graph corresponding to the historical generated content is determined. The risk of cognitive homogenization of the differentiated content is determined based on the similarity.

4. The method as described in claim 2, characterized in that, The step of assigning weights to semantic elements in the semantic element set associated with the specific regional term based on the usage frequency information of the specific regional term includes: The breadth of geographical coverage is determined based on the number of administrative regions covered by the industry-specific entity set and the frequency of use of the specific regional terms. Determine the semantic association between multiple words in the original corpus and the industry-specific entity set; The conflict index of multiple words in the original corpus is determined based on the semantic relevance. The importance of each of the plurality of words is determined based on the geographical coverage and the conflict index.

5. The method as described in claim 4, characterized in that, After determining the importance of each word among the plurality of words based on the geographical coverage and the conflict index, the method further includes: The original corpus was sequentially input into a religiously sensitive word database and a regional folklore word database for filtering. Perform a substitution word mapping transformation on the filtered words to obtain a set of semantic elements and the importance of the words in the set of semantic elements.

6. The method as described in claim 5, characterized in that, The determination of the information entropy of the industry-specific entity set includes: Determine the entity distribution status of the industry-specific entity set: The percentage of different types of entities in the industry-specific entity cluster is statistically analyzed. The information entropy of the industry-specific entity set is determined based on the proportion of different types of entities. A higher entropy value indicates a more discrete distribution of entity categories in the industry-specific entity set, and a higher degree of dispersion in the distribution of entity categories in the industry-specific entity set indicates stronger diversity of industry knowledge.

7. The method as described in claim 6, characterized in that, The step of configuring a dynamic sampling temperature function that fluctuates with the generation time period based on the information entropy includes: The fluctuation amplitude coefficient and base value parameter of the preset temperature function are adjusted according to the entropy value of the information entropy to obtain the dynamic sampling temperature function. The fluctuation amplitude coefficient is directly proportional to the information entropy value, and the base value parameter is inversely proportional to the information entropy value.

Images