Interaction scheme generation method and apparatus, electronic device, and storage medium

Abstract

The application relates to the technical field of interaction scheme generation, and discloses an interaction scheme generation method and device, electronic equipment and a storage medium, wherein the method comprises the following steps: converting abstract cultural dimensions into specific interaction control parameters by constructing a quantifiable cultural adaptation rule library; fusing and analyzing structured and unstructured business data through a multi-modal data processing model to generate a target feature vector representing an interaction object, providing an integrated basis for model decision-making, and enhancing the individuality and rationality of the scheme; and further combining the interaction control parameters and the target feature vector into a structural control condition to guide a large language model, combining domain knowledge with model generation capability. The application has the beneficial effect that the automation and intelligence of the interaction scheme generation process are realized, thereby improving the strategy generation efficiency and consistency in a cross-cultural business interaction scenario.

Description

Technical Field

[0001] This invention relates to the field of interactive scheme generation technology, and in particular to an interactive scheme generation method, apparatus, electronic device and storage medium. Background Technology

[0002] In a globalized business environment, effective communication and negotiation with stakeholders from different cultural regions has become the norm. Current technologies heavily rely on human experience in formulating interaction plans, making automatic generation difficult. Firstly, cross-cultural differences are hard to quantify and integrate into the decision-making process, and human experience is difficult to systematize, resulting in poor cultural adaptability of the generated plans, easily leading to misunderstandings or conflicts. Secondly, the information upon which decision-making depends is fragmented; multi-source heterogeneous information such as business data (e.g., transaction records) and text data from the interaction stakeholders are isolated, lacking effective means of fusion and analysis, making it difficult to form a comprehensive and accurate profile of the interaction stakeholders. Summary of the Invention

[0003] Based on this, it is necessary to propose an interaction scheme generation method, device, electronic device and storage medium to address the existing problem of interaction scheme generation.

[0004] An interaction scheme generation method, the method comprising:

[0005] Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0006] Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base;

[0007] The target rule set is dimensionalized to obtain the interaction control parameters;

[0008] The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0009] The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model;

[0010] Based on the control conditions, an interaction scheme is generated using the large language model.

[0011] Further, the step of dimensionalizing the target rule set to obtain the interaction control parameters includes:

[0012] Extract the cultural dimension theoretical indicators from the target rule set, and the weight coefficients corresponding to each cultural dimension theoretical indicator;

[0013] The theoretical indicators of each cultural dimension and their corresponding weight coefficients are mapped to the interactive control parameters according to a preset mapping method.

[0014] Furthermore, the step of fusing and analyzing the multi-source heterogeneous data using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object includes:

[0015] The unstructured business data in the multi-source heterogeneous data is parsed using a natural language processing model to obtain semantic features, and the structured business data in the multi-source heterogeneous data is extracted to obtain quantitative indicator data.

[0016] The semantic features are associated and aligned with the quantitative indicator data;

[0017] The target feature vector is generated based on the semantic features after association alignment and the quantitative index data.

[0018] Furthermore, after the step of generating an interaction scheme based on the control conditions and the large language model, the method further includes:

[0019] Obtain the compliance constraint rules from the target rule set;

[0020] The compliance constraint rules are used to detect whether the interactive content in the interaction scheme is compliant.

[0021] If the interactive content in the interaction scheme is compliant, the interaction scheme will be sent to the designated terminal.

[0022] Furthermore, after the step of generating an interaction scheme based on the control conditions and the large language model, the method further includes:

[0023] Record the process data and result data of the interaction with the specified interaction object based on the interaction scheme;

[0024] Based on the process data and the result data, the target rule set corresponding to the culture adaptation rule base is updated and optimized.

[0025] Furthermore, before the step of calling the corresponding target rule set from the pre-built cultural adaptation rule base based on the region identification information, the method further includes:

[0026] Obtain original cultural characteristic data for each cultural region;

[0027] Based on the preset cultural dimension quantification model, the original cultural feature data of each region are processed to generate a basic rule set corresponding to each cultural region.

[0028] Acquire multiple interaction scenarios and successful interaction case samples for each scenario;

[0029] Analyze the successful interaction case samples and extract strategy features related to the quantitative cultural dimension in the basic rule set;

[0030] The strategy features are bound to the corresponding cultural regions, interaction scenarios, and quantified cultural dimensions to form rule entries;

[0031] All rule entries are compiled to build the culture-adaptive rule base.

[0032] Furthermore, the step of generating an interaction scheme based on the control conditions using the large language model includes:

[0033] Based on the control conditions, multiple candidate interaction schemes are generated using the large language model;

[0034] Based on a preset set of evaluation rules, each candidate interaction scheme is simulated and evaluated to predict the expected interaction effect of each candidate interaction scheme.

[0035] Based on the expected interaction effect of each candidate interaction scheme, the target interaction scheme is selected from the multiple candidate interaction schemes as the output interaction scheme.

[0036] An interactive scheme generation apparatus, the apparatus comprising:

[0037] The acquisition module is used to acquire the region identification information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0038] The calling module is used to call the corresponding target rule set from the pre-built culture adaptation rule base based on the region identification information;

[0039] The quantization module is used to quantize the dimensions of the target rule set to obtain interactive control parameters;

[0040] The analysis module is used to perform fusion analysis on the multi-source heterogeneous data through a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0041] The input module is used to input the interactive control parameters and the target feature vector as control conditions into the pre-trained large language model;

[0042] The generation module is used to generate an interaction scheme based on the control conditions and the large language model.

[0043] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0044] Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0045] Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base;

[0046] The target rule set is dimensionalized to obtain the interaction control parameters;

[0047] The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0048] The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model;

[0049] Based on the control conditions, an interaction scheme is generated using the large language model.

[0050] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0051] Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0052] Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base;

[0053] The target rule set is dimensionalized to obtain the interaction control parameters;

[0054] The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0055] The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model;

[0056] Based on the control conditions, an interaction scheme is generated using the large language model.

[0057] The beneficial effects of this invention are as follows: By constructing a quantifiable cultural adaptation rule base, abstract cultural dimensions are transformed into specific interaction control parameters, providing cultural constraints for the large language model and improving the cultural adaptability and regional compliance of the generated solutions; simultaneously, by fusing and analyzing structured and unstructured business data through a multimodal data processing model, target feature vectors representing interactive objects are generated, providing an integrated basis for model decision-making and enhancing the individual relevance and rationality of the solutions; furthermore, by combining interaction control parameters and target feature vectors into structured control conditions to guide the large language model, domain knowledge is combined with model generation capabilities, realizing the automation and intelligence of the interaction solution generation process, thereby improving the efficiency and consistency of strategy generation in cross-cultural business interaction scenarios. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] in:

[0060] Figure 1 This is an application environment diagram of the interaction scheme generation method in one embodiment;

[0061] Figure 2 Here is a flowchart of an interaction scheme generation method in one embodiment;

[0062] Figure 3 This is a structural block diagram of an interaction scheme generation device in one embodiment;

[0063] Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Figure 1 Generate an application environment diagram for the interaction scheme in one embodiment. (Refer to...) Figure 1This interaction scheme generation method is applied to an interaction scheme generation system. The interaction scheme generation system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet computer, or laptop computer. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to obtain the regional identification information of a specified interactive object, and the server 120 is used to generate the interaction scheme.

[0066] like Figure 2 As shown, in one embodiment, an interaction scheme generation method is provided. This method can be applied to both terminals and servers; this embodiment uses terminal application as an example. The interaction scheme generation method specifically includes the following steps:

[0067] S1: Obtain the region identifier information of the specified interactive object, and the multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0068] S2: Based on the regional identification information, call the corresponding target rule set from the pre-built cultural adaptation rule base;

[0069] S3: Perform dimensionality quantization on the target rule set to obtain interactive control parameters;

[0070] S4: The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0071] S5: Input the interactive control parameters and the target feature vector as control conditions into the pre-trained large language model;

[0072] S6: Based on the control conditions, generate an interaction scheme through the large language model.

[0073] As described in step S1 above, obtain the regional identification information of the specified interactive object, as well as the multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data. First, identify the regional identification information of the interactive object, which may come from supplier profiles, IP addresses, company registered addresses, country / region fields in transaction contracts, or geocodes entered by users. After identification, it needs to be standardized, and the regional granularity (country, province / state, city, or industry segmentation) needs to be recorded. Then, collect multi-source heterogeneous data. Examples of structured data include historical transaction records exported from ERP / CRM (order volume, on-time delivery rate, number of overdue payments, etc.), financial statement indicators (revenue, accounts receivable turnover), and contract term indexes; examples of unstructured data include supplier website text, news reports, social media posts (multilingual), email conversations, and scanned copies of compliance documents. Technically, a data collector needs to be configured: data is crawled through APIs, database connectors, secure crawlers (the secure crawlers are configured to comply with the robots.txt protocol of the target website and the data regulations of the region), and document ingestion modules (which can be optical character recognition technology). The raw data is then preprocessed: timestamp standardization, deduplication, outlier detection, language recognition and encoding unification; sensitive information is anonymized and audited for compliance, and the source, crawling time and confidence level are recorded in the metadata. The output is a structured raw dataset merged with metadata index for subsequent feature extraction and profile construction.

[0074] As described in step S2 above, the corresponding target rule set is retrieved from the pre-built cultural adaptation rule base based on the regional identification information. Relying on the pre-built cultural adaptation rule base (a searchable knowledge base), entries are divided by region / industry / scenario. Each record contains cultural dimension indicators, applicable scenario tags, priority, version number, and adaptation meta-rules. The retrieval process includes: determining candidate rule sets based on standardized regional identifiers (supporting hierarchical fallback, e.g., city → country → region), and then performing secondary filtering according to interaction scenario tags (e.g., initial inquiry, renewal negotiation, urgent delivery). Retrieval can employ full-text indexing + attribute filtering, or semantic matching based on ontology, to handle similar regions or cases lacking explicit mappings. The selected rule set is returned in a machine-readable format (e.g., JSON), containing the cultural dimensions to be quantified (e.g., communication directness) and initial metadata (weight suggestions, confidence levels). In practice, version control, caching mechanisms, and audit logs for the rule base need to be implemented to ensure that subsequent rule updates and backtracking are traceable.

[0075] As described in step S3 above, the target rule set is dimensionalized to obtain interaction control parameters. The descriptive cultural dimensions in the rule set are mapped into computable interaction control parameters. First, a basic value is assigned to or read for each cultural dimension (which can be derived from existing indices such as Hofstede or calibrated using historical data). The value is then normalized (e.g., mapped to the [0,1] interval). Next, the normalized dimensions are converted into specific control parameters according to the mapping function. For example, "communication directness" is mapped using a linear or sigmoid function to generate a "bargaining directness coefficient"; "relationship orientation" generates a "relationship maintenance weight"; and "time perception" is mapped to a "response time limit priority". The mapping function can be a linear weighted model, a piecewise function, or a learning mapping based on a regression / classification model (when historical cases are used for supervised training). To reflect uncertainty, the parameter confidence interval or confidence score can be output simultaneously. The parameter set is stored in a structured format (name, value, unit, source, confidence level, expiration date) and can be directly read by downstream modules as a constraint or priority input. It can also include industry / scenario correction factors (for example, the bargaining threshold in the automotive parts industry is different from that in the clothing industry), and the baseline parameters are adjusted by multiplication and division factors.

[0076] As described in step S4 above, the multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object. Semantic and numerical fusion is performed on structured and unstructured data to produce a unified target feature vector. The processing flow includes several sub-modules: text preprocessing (word segmentation, stop word filtering, language detection, entity recognition), visual / document processing (OCR, layout parsing, table recognition), and numerical indicator processing (time series summarization, statistical calculation). Unstructured text is encoded into text embeddings using cross-lingual pre-trained models (such as XLM-R or multilingual BERT). Structured numerical values ​​are standardized and feature-engineered before being mapped to numerical vectors. A fusion strategy is then employed: simple concatenation, weighted averaging, or attention-based cross-modal fusion networks (such as Transformer fusion layers) to integrate the embeddings from different modalities. When it is necessary to associate entities with historical relationships, knowledge graph construction and graph neural networks (GNNs) can be used for topological information fusion. After fusion, high-dimensional vectors can be reduced in dimensionality or made sparse (PCA, Autoencoder), and the vectors are normalized for subsequent model consumption. Simultaneously, metadata corresponding to the vectors (time window, data coverage, missing rate, confidence level) is generated. The output vectors should encode information including: fulfillment capability, financial health, cooperation intention signals (based on sentiment / semantic analysis), and supply capacity risk indicators, which can be used as control conditions input to downstream models.

[0077] As described in step S5 above, the interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model. The key is to effectively convert the numerical control parameters and high-dimensional feature vectors into a form that the large language model (LLM) can accept and utilize. A common approach is to construct a structured prompt or provide context to the LLM using function calls / tool ​​interfaces. The specific process is as follows: The interactive control parameters are formatted into constraint text segments according to a predefined template (e.g., "negotiation directness coefficient = 0.7, first-round concession limit = 10%"); key features are extracted and textualized from the feature vector (summary description or mapping the vector back to explanatory labels, such as "high performance history / recent posts indicate tight capacity"), and metadata (confidence level, data source) is attached. The above text segments are concatenated with task instructions, role settings (e.g., "You are a cross-border procurement strategy generator") and generation format specifications (JSON schema or action sequence template) to form structured prompt words. If the LLM supports tool calls or function APIs, numerical parameters can be passed in as machine-readable parameters, and the LLM will trigger different generation logic based on these parameters. In addition, to control the context window, a search-enhanced generation (RAG) mechanism can be used to inject necessary document fragments into the search engine as prompts, including input validation and security filtering (blocking sensitive words and compliance restrictions) to prevent non-compliant output.

[0078] As described in step S6 above, an interaction scheme is generated through the large language model based on the control conditions. Upon receiving structured prompts or parameterized input, the LLM generates and outputs the interaction scheme. To ensure the executability and parsability of the scheme, the output typically uses a predefined structured format (such as JSON), including: step-by-step interaction actions (sequential steps), triggering conditions for each step, parameter values ​​(such as target discount range, time limit), alternative branch logic, and compliance prompts. The generation process can adopt a multi-candidate generation strategy: the model produces several candidate schemes, which are then scored, simulated, or validated by a subsequent evaluation module before being output as the best. The generated results require post-processing: semantic validation (whether hard constraints are met), format validation (whether it conforms to the JSON schema), sensitive word replacement, and final compliance scanning. The output can then be presented to end users for human review or converted into machine-executable instructions (to ensure the executability and parseability of the solution, the output typically uses a predefined structured format (such as JSON), including: step-by-step interactive actions (sequential steps), triggering conditions for each step, parameter values ​​(such as target discount range, time limit), backup branch logic, and compliance prompts) and trigger external system calls. It is recommended to record the generated version, prompt snapshots, and model responses throughout the entire process for interpretability tracking and subsequent online learning or rule base updates.

[0079] In one embodiment, step S3, which involves dimensionalizing the target rule set to obtain the interaction control parameters, includes:

[0080] S301: Extract the cultural dimension theoretical indicators from the target rule set, and the weight coefficients corresponding to each cultural dimension theoretical indicator;

[0081] S302: Map each of the cultural dimension theoretical indicators and the corresponding weight coefficients to the interactive control parameters according to a preset mapping method.

[0082] As described in step S301 above, the theoretical indicators of cultural dimensions in the target rule set, and the corresponding weight coefficients for each cultural dimension theoretical indicator, are extracted. First, entries related to cultural dimensions are retrieved from the called target rule set (usually a machine-readable structured format, such as JSON, YAML, or a database table). Each entry contains a cultural dimension identifier (e.g., "Uncertainty Avoidance (UAI)"), dimension source (e.g., Hofstede index, regional survey data, or expert assignment), original measurement value, and initial weight coefficient fields. In implementation, the rule set needs to be syntax- and semantically validated: confirming that the identifier for each dimension is valid, the original measurement value is within the expected range, and the weight coefficient is a numerical value within [0,1] or a specified range. For cases with hierarchical coverage (e.g., city rules covering national rules), entries are merged according to priority (scenario priority, region priority, or explicit version number), and the source is recorded. If some commonly used dimensions are missing from the rule set, a fallback strategy is triggered (using the average value of similar regions or the default library value), and the imputation source is recorded in the metadata. The engineering implementation suggests temporarily storing the extracted results in a structured object (e.g., rules_features[]), containing the fields: dimension_id, raw_value, normalized_hint, weight_coefficient, source, and confidence. For ease of subsequent traceability, the rule set version number, extraction timestamp, and unique identifier (URI or primary key) of each rule entry must be recorded simultaneously. Furthermore, the weight coefficients should be validated within a reasonable range, and anomaly alerts should be issued (e.g., prompting manual review when the total weight exceeds a threshold).

[0083] As described in step S302 above, the theoretical indicators of each cultural dimension and their corresponding weight coefficients are mapped to the interaction control parameters according to a preset mapping method. After the extraction of dimensions and weights is completed, the descriptive cultural indicators are transformed into numerical parameters that can be directly used to control interactive behavior. First, the original dimension values ​​are standardized (Min-Max normalization or Z-score standardization can be selected) to eliminate the differences in the scale of different dimensions. Then, according to a preset family of mapping functions (e.g., linear mapping, Sigmoid / Logistic mapping, piecewise mapping, or regression models based on supervised learning), the standardized dimensions are mapped to one or more interaction control parameters (e.g., negotiation directness coefficient, relationship maintenance weight, risk clause detail, response time priority, etc.). Example mapping: Negotiation directness = α * (1 - norm(PD)) + β * norm(IDV), where α and β are mapping weights, and norm() represents the normalized dimension value; or a trained model can be used: param = f_theta([norm(PD), norm(UAI), …]), where f_theta is a regression or neural network model trained based on historical interaction results. Scenario / industry correction factors (read from rule entries) should also be applied during the mapping process, such as multiplying the "first inquiry" scenario by coefficient γ1 and the "urgent delivery" scenario by γ2 to achieve scenario adaptation. To reflect uncertainty, the mapping module should calculate and output the confidence level or interval of the parameters (e.g., through Monte Carlo sampling or based on model prediction variance estimation). Finally, boundary constraints and quantization are applied to the generated interactive control parameters (e.g., limited to [0,1] or discrete levels 1-5), and the results are output in a structured format (fields include parameter_id, value, unit, confidence, derived_from_dimensions, mapping_function_id, and effective_timestamp). In engineering practice, configurable mapping functions (supporting hot updates and A / B testing), a training / validation pipeline for the mapping model, and logging and version control of the mapping results are required to facilitate subsequent backtracking, evaluation, and iterative optimization. For missing or low-confidence data, a degradation strategy (using regional averages or historical good values) should be triggered, and the reason for degradation should be recorded.

[0084] In one embodiment, step S4, which involves fusing and analyzing the multi-source heterogeneous data using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object, includes:

[0085] S401: Use a natural language processing model to parse the unstructured business data in the multi-source heterogeneous data to obtain semantic features, and extract the structured business data in the multi-source heterogeneous data to obtain quantitative indicator data;

[0086] S402: Associate and align the semantic features with the quantitative indicator data;

[0087] S403: Generate the target feature vector based on the semantic features after association alignment and the quantitative index data.

[0088] As described in step S401 above, the unstructured business data in the multi-source heterogeneous data is parsed using a natural language processing model to obtain semantic features, and the structured business data in the multi-source heterogeneous data is extracted to obtain quantitative indicator data. First, preprocessing is performed on the unstructured business data: language detection and standardization, character encoding unification, noise removal, and, if necessary, OCR (text extraction from scanned documents or images). Then, word segmentation / sub-wording is performed. For multilingual text, it is recommended to use cross-language pre-trained models (such as XLM-RoBERTa, mBERT, etc.) or models fine-tuned in the target domain to obtain context-sensitive text embeddings. Based on these embeddings, further named entity recognition (NER), relation extraction, topic / intent detection, sentiment / stance analysis, and event extraction (such as delays, complaints, and cooperation intention handles) can be performed, and the extracted semantic tags, timestamps, main entities, and signal strengths are output as semantic features. For documents involving tables or contract terms, a table understanding / contract parser can be used to extract key items (such as payment terms, delivery dates, and default clauses). For structured business data (such as ERP / CRM exported tables, financial indicators, and transaction records), the following steps are performed: field validation, time series aggregation (calculating the mean / year-on-year / month-on-month by daily / weekly / monthly windows), missing value completion and normalization, and extraction of key information items (on-time delivery rate, number of overdue orders, average order amount, etc.) to form quantitative indicator data. The entire process should output a semantic feature set with metadata and a quantitative indicator table, including source, time, confidence level, and processing link, so as to trace the source and assign weights during subsequent alignment and fusion.

[0089] As described in step S402 above, the semantic features are associated and aligned with the quantitative indicator data. The goal is to establish a usable correspondence between the semantic signals in the text and the structured numerical indicators in the entity and time dimensions. First, entity resolution is performed: the company / product / contract entities identified in the text are matched with unique identifiers in the structured records (such as company tax number, contract number, SKU) using string similarity, rule matching, embedding-based semantic matching, or knowledge graph linking methods to generate entity alignment mappings. Second, time alignment is performed: the timestamps of text events are matched with the time series windows of structured data (e.g., aligning "tight production capacity in the past three months" with the shipment / delay rate sequence of the past 90 days), and time weights are calculated (more recent information is assigned higher weights). The semantic-to-quantitative mapping can use rule mapping (e.g., mapping "high willingness to cooperate" to a willingness to cooperate score + 0.8) or learning-based methods (training a small regression / classifier to predict text features as correction terms for numerical indicators). To handle multiple conflicting information, conflict resolution strategies should be designed (weighting by source confidence, merging by time weight, or triggering manual verification). Technically, a knowledge graph can be used to represent entity relationships, events, and numerical indicators as a graph structure, and then graph matching or graph neural networks (GNNs) can be used for cross-modal information fusion and alignment. The output is an aligned feature table: the mapping relationship between the semantic features of each entity and the corresponding quantitative indicator under each time window, along with confidence level, source, and alignment method identifiers.

[0090] As described in step S403 above, the target feature vector is generated based on the semantic features after association alignment and the quantitative index data. After obtaining the aligned multimodal input, feature fusion is performed to generate a unified target feature vector. Common implementations include: first, performing dimensionality transformation on each modality feature (projecting different modalities to the same dimension through linear mapping or a small feedforward network), and then using fusion mechanisms such as concatenation, weighted averaging, or attention-based cross-modal fusion layers (multi-head attention) to automatically learn the weight allocation of each feature. For entity relationship information, graph structure features can be input into a GNN and connected with the representation vector to introduce topological information. After fusion, the high-dimensional vector can be regularized and dimensionality reduced (e.g., BatchNorm, PCA, Autoencoder), and interpretability labels for each sub-dimension (e.g., "performance capability component", "financial health component", "cooperation intention component", "compliance risk component") can be calculated. These components can be output by explicit aggregation functions (weighted sum, percentile score) or by downstream evaluation models. In addition, metadata should be output: data coverage, missing item markers, overall confidence level, and generation timestamp. The final target feature vector should be used as input to the LLM as subsequent control conditions, and should also be stored in a machine-readable format (e.g., in a vector database or key-value pairs), exposing an API for policy evaluation, simulation, or online learning modules. Engineering considerations include real-time / offline generation modes, incremental update strategies and batch throughput optimization, as well as degradation strategies for low-confidence or missing inputs (using historical averages or default vectors).

[0091] In one embodiment, after step S6 of generating an interaction scheme based on the control conditions using the large language model, the method further includes:

[0092] S701: Obtain the compliance constraint rules in the target rule set;

[0093] S702: Detect whether the interactive content in the interaction scheme is compliant based on the aforementioned compliance constraint rules;

[0094] S703: If the interaction content in the interaction scheme is compliant, then the interaction scheme is sent to the designated terminal.

[0095] As described in step S701 above, the compliance constraint rules in the target rule set are obtained. First, compliance-related entries are located from the previously invoked target rule set. The rule set is typically stored in a structured format (e.g., JSON, YAML, or relational / document database records). Each compliance rule includes a rule identifier, applicable region / scenario tag, rule type (legal constraint, industry standard, cultural taboo, wording replacement rule, etc.), specific matching pattern (regular expression or semantic matching template), priority, and version information. During implementation, rule retrieval and filtering should be performed: applicable rules are filtered based on the current interaction scenario tag and region identifier, and sorted by priority and rule version. If the rule set supports hierarchical coverage (e.g., national laws take precedence over industry guidelines), the retrieval logic needs to implement hierarchical fallback and conflict resolution. Furthermore, the retrieved rules need to be semantically standardized, converting the natural language expressions in the rules into machine-executable detection specifications (e.g., converting "avoid involving interest" into matching keywords such as "interest" and "interest" and including a list of synonyms), and generating a rule table for the detector (field examples: rule_id, pattern_type, pattern_value, severity, action_hint, source_document). The implementation should record the rule source, retrieval timestamp, and applicability confidence level to facilitate subsequent auditing and rule update retrospectives.

[0096] As described in step S702 above, the compliance constraint rules are used to detect whether the interactive content in the interaction scheme is compliant. The detection process first parses the interaction scheme into structured units (e.g., step sequences, text fields, numerical parameters) according to a predefined format, and attaches contextual metadata (target object, scenario, triggering conditions) to each unit. The detection adopts a multi-level approach: First, hard rule detection based on pattern matching, using regular expressions, dictionary matching, and word substitution tables to verify each text field, applicable to explicit prohibited words, legal terms, or contractual clause formats; second, soft rule detection based on semantics, using pre-trained language models or fine-tuned classifiers to perform sensitivity discrimination, intent recognition, and meaning recognition on the text, capable of handling implicit expressions, synonym substitutions, and contextual meaning (e.g., determining whether a sentence contains implicit interest clauses); third, rule detection based on numerical thresholds, comparing the quantitative parameters (e.g., discount rate, delivery commitment) in the interaction scheme against compliance upper and lower limits. The detection results generate a judgment item (pass / fail / uncertain), violation type, violation location, severity score, and suggested action (replacement, deletion, manual review) for each rule. To reduce false alarms, the system can integrate evidence from multiple sources and weight it with confidence levels, and trigger manual review or enable automatic correction processes for uncertainties. The entire detection module must output an auditable detection report (JSON format), including rule references, matching fragments, confidence levels, and processing suggestions, for subsequent recording and traceability.

[0097] As described in step S703 above, if the interactive content in the interaction scheme is compliant, the interaction scheme is sent to the designated terminal. When the detection module returns that all necessary items are compliant (or passes the acceptable partial compliance threshold according to the preset strategy), the scheme distribution process is executed. First, the interaction scheme is formatted according to the interface specifications of the target terminal: if it is a manual terminal, a display document / message (such as an HTML / Markdown report or in-application card) is generated; if it is an automated execution terminal, it is converted into a machine-executable instruction set (such as JSON instructions, API call sequences, or message queue tasks), and accompanied by necessary metadata (scheme version, rule version, generation time, confidence level, audit ID). Then, the data is sent to the designated terminal through a secure channel: for external services, it is published through an HTTPS interface or message queue; for local terminals, it is published through local IPC or mobile push. At the same time, access control and logging are implemented to record the sending operator / system account, target terminal identifier, and response status. If the scheme involves sensitive information, encryption and desensitization strategies must be implemented before transmission and data specification requirements must be followed. If the detection result is partially compliant or non-compliant, the system should execute an alternative path according to the policy: automatically generate a revised compliant version and re-detect, notify for manual review, or save the solution as a draft and record the rejection reason. After distribution, the system should receive and record the acknowledgments or execution results from the target terminals to drive subsequent feedback updates and rule iterations.

[0098] In one embodiment, after step S6 of generating an interaction scheme based on the control conditions using the large language model, the method further includes:

[0099] S711: Record the process data and result data of interacting with the specified interaction object based on the interaction scheme;

[0100] S712: Based on the process data and the result data, update and optimize the target rule set corresponding to the culture adaptation rule base.

[0101] As described in step S711 above, record the process data and result data of the interaction with the specified interaction object based on the interaction scheme. The entire interaction process should be recorded in a structured manner to support subsequent analysis and rule base optimization. Recorded content includes, but is not limited to: interaction scheme identifier (scheme version, generation time, corresponding rule set version number, prompt word snapshot), interaction trigger conditions, sending / execution timestamp, target interaction object identifier, input / output of each interaction action (e.g., sent text content, API call parameters, received reply text or status code), interaction duration, response delay, interaction interruption or abnormal events and their error codes, and final interaction result (e.g., the other party accepts / rejects / negotiates subsequent terms, whether the order is completed, negotiated discount amount). In addition, contextual metadata should be collected synchronously, such as interaction channels (email, platform messages, API), participant accounts, geographical / time zone information, and compliance review records (detection pass / fail and modification records). For engineering implementation, it is recommended to adopt an event-driven logging system (such as Kafka or other message queues), writing interaction logs to persistent storage (relational database or log storage system) and synchronizing them to the analysis repository; for large-scale vector or multimedia data, references to object storage can be written to the meta table. Records must include data quality tags (missing rate, confidence level) and compliance markers (whether it contains sensitive information, whether it has been anonymized); all records must have tamper-proof audit information (operator / system, timestamp, signature) for traceability and legal compliance.

[0102] As described in step S712 above, the target rule set in the cultural adaptation rule base is updated and optimized based on the process data and the result data. This process revolves around the goal of "closed-loop rule optimization from operational data," encompassing four sub-processes: data preprocessing, effect evaluation, parameter adjustment, and release. First, the recorded data is cleaned and labeled: noise reduction, error checking, and missing value handling are performed. Interaction results are then mapped to quantitative evaluation indicators (such as acceptance rate, conversion rate, average discount, interaction rounds, compliance violation rate, etc.). Second, causal or statistical analysis is conducted: the impact of a rule or parameter on business indicators is evaluated through group comparison, regression analysis, or causal inference methods. If necessary, A / B testing or multi-option optimization models are performed to verify the effectiveness of candidate adjustments. Based on the analysis results, rule adjustment suggestions are generated. Adjustment methods include directly modifying the weight coefficients in the rules, adding or deleting scenario correction factors, updating the mapping function, or training a new mapping model (supervised learning or online learning). For model-based mapping, a training-validation-test pipeline should be constructed, using the most recent interaction data for incremental training, and its effectiveness and security are verified through offline simulation and online canary release. Rule changes must undergo automated regression testing and compliance checks. Each change generates a new rule version and records the reason for the change, rollback point, and verification results. High-risk modifications should trigger a manual review process. Finally, the verified rules or models are published to the rule repository, triggering downstream cache refresh. Key metrics are continuously monitored; if a metric regresses, the system automatically rolls back to the previous version. The entire optimization process should maintain a complete audit trail and interpretable reports for monitoring, review, and continuous iteration.

[0103] In one embodiment, before step S2, which involves calling the corresponding target rule set from a pre-built culture-adaptive rule base based on the region identification information, the method further includes:

[0104] S101: Obtain original cultural characteristic data for each cultural region;

[0105] S102: Based on the preset cultural dimension quantification model, process each of the original cultural feature data to generate a basic rule set corresponding to each cultural region;

[0106] S103: Obtain multiple interaction scenarios and a sample of successful interaction cases for each scenario;

[0107] S104: Analyze the successful interaction case samples and extract the strategy features related to the quantitative cultural dimension in the basic rule set;

[0108] S105: Bind the strategy features with the corresponding cultural regions, interaction scenarios, and quantified cultural dimensions to form rule entries;

[0109] S106: Gather all rule entries and construct the culture-adaptive rule base.

[0110] As described in step S101 above, original cultural characteristic data for each cultural region is acquired. This involves collecting original characteristic data describing each cultural region through multiple channels. This data includes both structured statistical data and unstructured text and symbolic materials. Structured data sources may include publicly available cultural index databases (such as the Hofstede Index, World Values ​​Survey), and socioeconomic indicators (education level, cultural composition, language distribution, internet penetration rate, etc.) published by national / regional statistical bureaus. Unstructured data sources may include academic papers, industry reports, local news, local regulations, and social media discussion corpora. Technically, a data acquisition pipeline needs to be established: open data is acquired via API or batch download; documents and reports are acquired via web crawlers or content subscription systems (complying with copyright and usage licenses); and social media corpora are acquired using authorized acquisition interfaces. Metadata (source, time, credibility, licensing information) is recorded simultaneously during the acquisition process. Sensitive or restricted data must undergo compliance review beforehand and be anonymized or aggregated as necessary. The output is the raw feature dataset indexed by cultural region. It is recommended to save it in a standardized storage format (such as JSON-Lines or tables with region as the primary key) and label the data version and source in the storage for subsequent traceability and updates.

[0111] As described in step S102 above, based on the preset cultural dimension quantification model, the original cultural feature data for each region are processed to generate a basic rule set corresponding to each cultural region. First, the model framework for quantifying cultural features is determined, for example, using Hofstede's six dimensions or a custom dimension system as a benchmark, and defining the measurement indicators and normalization methods for each dimension. The cultural dimension quantification model, based on a predefined cultural dimension system (such as Hofstede's cultural dimension theory), analyzes the text content in the original cultural feature data using natural language processing techniques such as keyword extraction, sentiment analysis, and topic modeling. Combined with structured statistical data, it outputs the quantified scores for each dimension through weighted or model-based calculations. The original data for each region is cleaned and standardized (missing data imputation, anomaly detection, unit unification), and the original scores are calculated according to the dimension definitions, such as combining multiple indicators into a single dimension score through weighted averaging. To ensure comparability, the scores are normalized (e.g., Min-Max or Z-score), and the normalization parameters are recorded. The normalized dimension scores and confidence scores (calculated based on data coverage and source credibility) are then used to construct basic rule units. Each unit includes the following fields: region_id, dimension_id, raw_score, normalized_score, confidence, source_list, and timestamp. To accommodate applications with different granularities, the basic rule set should support multiple levels (national, provincial, and city levels) and implement fallback logic (using the national or regional average when city-level data is insufficient). Furthermore, it is recommended to introduce an expert verification process: sample the quantification results for review and adjust weights or mapping rules based on expert feedback, outputting a verified basic rule set and incorporating it into version control and audit records.

[0112] As described in step S103 above, multiple interaction scenarios and successful interaction case samples for each scenario are obtained. The aim is to define a set of representative interaction scenarios (e.g., initial inquiry, contract renewal, urgent delivery request, price negotiation, compliance communication, etc.) and collect and label successful interaction case samples for each scenario. Scenario definitions need to clearly define triggering conditions, participant roles, target metrics (e.g., conversion rate, response time, discount rate, etc.), and quantitative standards for evaluating success. Case samples can be obtained from enterprise historical interaction logs (ERP / CRM, email records, negotiation records), public case libraries, or anonymized data provided by partners. Compliance reviews are required when collecting samples; de-identification or aggregation may be necessary to protect sensitive information. The collected cases are preprocessed: categorized by scenario, segmented by time window, key fields extracted (interaction text, timestamp, result tags, relevant numerical indicators), and metadata synchronized (region, industry, participant type). To ensure the reliability of subsequent analysis, the samples should be screened for quality (incomplete or noisy samples should be removed) and labeled manually or semi-automatically (success / failure labels, key strategy fragment labels) to form a high-quality case set for analysis. This set should be saved as an indexable sample library (such as a labeled JSON file or database table) and the sample source and authorization information should be recorded.

[0113] As described in step S104 above, the successful interaction case samples are analyzed to extract strategy features related to the quantitative cultural dimension in the basic rule set. A combination of data-driven and rule-driven methods is used to extract strategy features that can be mapped to the cultural dimension from the successful cases. First, the text of the samples undergoes NLP processing (word segmentation, entity recognition, event extraction, intent recognition, sentiment analysis), and numerical information is statistically summarized (e.g., average discount, number of negotiation rounds, delivery commitment duration). Then, based on statistical association analysis or causal exploration methods (correlation analysis, regression model, decision tree, or causal inference), the correlation between certain strategy features (e.g., asking price range, politeness of wording, and advance concession strategy) and success labels is evaluated, and it is determined how these features interact with the cultural dimension (e.g., strong relationship orientation). Clustering algorithms can also be used to identify common effective strategy patterns within the same cultural context, and the patterns are ranked using indicators such as frequency and success rate. The discovered features are normalized and structured (such as strategy feature ID, description, triggering conditions, typical value range, associated cultural dimensions, and significance indicators), and the key mappings are verified by domain experts or rule reviewers to form a set of strategy features that can be used for parameterization.

[0114] As described in step S105 above, the strategy features are bound to the corresponding cultural regions, interaction scenarios, and quantitative cultural dimensions to form rule entries. The aforementioned strategy features are structured and bound to cultural dimensions and scenario semantics to generate rule entries that can be retrieved and invoked by the system. Each rule entry should include a correlation key (region_id, scenario_id, dimension_id), a strategy feature description, suggested parameterized values ​​or parameter mapping functions (e.g., mapping "polite wording" to "relationship maintenance weight = 0.8" or providing specific wording templates), applicable conditions (e.g., object type, transaction size threshold), confidence level and source (data-driven or expert rule), and version and validity information. The binding process needs to implement priority and conflict strategies: if multiple rules conflict within the same region, they should be sorted according to rule priority, data confidence level, or recentity, while maintaining fallback logic. From an engineering perspective, rule entries are stored in a machine-readable format (e.g., JSON schema) and support indexed fields for fast retrieval (querying by region / scenario / dimension). In addition, it is recommended to save the reference relationship between the rule entries and the original samples together for subsequent auditing and interpretability analysis. Each time an entry is generated, it should be written to the change log and the data snapshot and verification results used during generation should be recorded.

[0115] As described in step S106 above, all rule entries are collected to construct the culture-adaptive rule base. All formed rule entries are summarized, validated, and organized into a workable culture-adaptive rule base. This base should implement a data model (entry table, index, version table), access interfaces (retrieval API, batch import / export interface), and management functions (rule editing, manual review, version release, rollback). During the construction process, consistency verification (checking duplicate entries, conflict detection, and coverage level legality), compliance review (ensuring no illegal, discriminatory, or sensitive content), and performance optimization (pre-calculating frequently used query caches, establishing regional / scenario combined indexes) must be completed. The rule base should support a hierarchical and inheritance mechanism (national level → industry level → scenario level) and provide rule merging strategies and conflict resolution mechanisms. Offline simulation testing is recommended before release: use historical interaction data replay to verify rule coverage and expected effects, and evaluate key indicators (such as suggested parameter distribution, coverage, and regression error). Once the rule base is built, it will be deployed to the production environment and monitored (access frequency, hit rate, error rate) and provided with operation and maintenance interfaces. At the same time, a rule update governance process (including automated updates, manual review channels, change auditing and rollback mechanisms) will be established to ensure that the rule base can continuously evolve and maintain traceability and controllability during operation.

[0116] In one embodiment, step S6, which generates an interaction scheme based on the control conditions using the large language model, includes:

[0117] S601: Based on the control conditions, generate multiple candidate interaction schemes using the large language model;

[0118] S602: Based on a preset set of evaluation rules, perform simulation evaluation on each of the candidate interaction schemes and predict the expected interaction effect corresponding to each of the candidate interaction schemes;

[0119] S603: Based on the expected interaction effect of each of the candidate interaction schemes, select the target interaction scheme from the multiple candidate interaction schemes as the output interaction scheme.

[0120] As described in step S601 above, multiple candidate interaction schemes are generated using the large language model based on the control conditions. Using structured prompt words (composed of interaction control parameters, target feature vectors, and task instructions) as input, a pre-trained large language model is invoked to generate several candidate outputs to ensure diversity and coverage. Various sampling or decoding strategies can be employed: such as temperature sampling, Top-k / Top-p (nucleus) sampling to increase diversity, or beam search in deterministic scenarios to stabilize quality. The system should predefine the number of candidates N (e.g., N=5~20) and configure decoding control parameters (temperature, top-p, beam size) to balance diversity and executability. Each candidate output is recommended to be generated in a structured format (e.g., JSON schema), including a step sequence, triggering conditions, quantization parameters (discount range, time limit), compliance prompts, and generation confidence or model output probability. During the generation process, prompt word snapshots, decoding parameters, and model versions should be recorded, and preliminary syntax and format checks should be performed on the candidates (whether they meet the JSON schema requirements, and whether required fields exist). To improve candidate quality, Retrieval Augmentation Generation (RAG) can be used to inject relevant document fragments into the context, or a lightweight classifier can be called in parallel after generation to filter out drafts that are obviously prohibited or conflict with control conditions. The final output is a set of candidate interaction schemes and their metadata, which can be used for subsequent simulation evaluation.

[0121] As described in step S602 above, each candidate interaction scheme is simulated and evaluated based on a preset evaluation rule set to predict the expected interaction effect corresponding to each candidate interaction scheme. The simulation evaluation module takes the candidate scheme as input and predicts its expected effect based on a hybrid evaluation pipeline driven by rules and data. The rule-driven part performs hard screening and scoring on the schemes according to the preset evaluation rule set (including compliance checks, logical consistency, resource constraints, scenario adaptability, etc.); the data-driven part uses a trained prediction model (such as acceptance probability model, transaction volume regression model, time-to-transaction prediction, discount impact model) to calculate the expected values ​​of several key business indicators for each candidate scheme. The evaluation may also include simulation based on user / supplier behavior models: constructing or calling the response model of the target object (a classifier or reinforcement learning agent trained based on historical interaction data), and performing Monte Carlo simulations on the candidate schemes on multiple possible feedback paths to estimate expected returns, risk distribution, and confidence intervals. Examples of evaluation metrics include: expected acceptance rate, expected transaction discount, expected number of interaction rounds, compliance risk score, time cost estimate, and expected net utility. To accommodate multiple dimensions, the evaluation output should include a point estimate, confidence interval, and evaluation rationale (citing the rules / models used) for each metric. The engineering implementation must support parallel evaluation, interpretable output (why this score was awarded), and the ability to trace the correlation between the evaluation results and the original candidates (saving evaluation versions, model versions, and input snapshots).

[0122] As described in step S603 above, based on the expected interaction effects of each candidate interaction scheme, a target interaction scheme is selected from multiple candidate interaction schemes as the output interaction scheme. The selection step implements multi-objective decision-making logic, transforming a set of indicators obtained from simulation evaluation into a comparable comprehensive score or selecting the Pareto optimal solution through multi-objective optimization. A common approach is to define a business-level weighted scoring function (weights can be dynamically adjusted by business strategy or online learning), normalize various criteria (e.g., acceptance rate, expected return, compliance risk, time cost), and then sum them by weight to obtain a comprehensive score, selecting the candidate with the highest score. Another approach is to first eliminate non-compliant or unqualified candidates based on threshold rules, and then use multi-objective ranking or Pareto frontier screening for the remaining candidates. For scenarios with uncertainty or risk sensitivity, risk-sensitive decision-making (e.g., maximizing CVaR, minimizing downside risk) or having the strategy output multiple alternatives for manual review in priority can be adopted. The selection process should retain interpretable records: explaining why a candidate was selected or rejected (referencing key evaluation indicators and rules), and writing this information into the audit log. Once the final target interaction scheme is determined, it can trigger subsequent compliance final review, be formatted into a terminal-usable format, and enter the sending / execution stage; at the same time, the eliminated candidates and their evaluation results are archived for subsequent rule / model training and A / B testing analysis.

[0123] Reference Figure 3 This solution also provides an interactive solution generation device, the device comprising:

[0124] The acquisition module 902 is used to acquire the region identification information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0125] The module 904 is used to call the corresponding target rule set from the pre-built culture adaptation rule base based on the region identification information;

[0126] The quantization module 906 is used to perform dimensionality quantization on the target rule set to obtain interactive control parameters;

[0127] Analysis module 908 is used to perform fusion analysis on the multi-source heterogeneous data through a preset multimodal data processing model to generate a target feature vector characterizing the specified interactive object;

[0128] Input module 910 is used to input the interactive control parameters and the target feature vector as control conditions into a pre-trained large language model;

[0129] The generation module 912 is used to generate an interaction scheme based on the control conditions and the large language model.

[0130] In one embodiment, the quantization module 906 includes:

[0131] The extraction submodule is used to extract the cultural dimension theoretical indicators from the target rule set, as well as the weight coefficients corresponding to each cultural dimension theoretical indicator.

[0132] The mapping submodule is used to map the theoretical indicators of each cultural dimension and the corresponding weight coefficients into the interactive control parameters according to a preset mapping method.

[0133] In one embodiment, the analysis module 908 includes:

[0134] The parsing submodule is used to parse the unstructured business data in the multi-source heterogeneous data using a natural language processing model to obtain semantic features, and to extract the structured business data in the multi-source heterogeneous data to obtain quantitative indicator data.

[0135] The alignment submodule is used to associate and align the semantic features with the quantitative indicator data;

[0136] The target feature vector generation submodule is used to generate the target feature vector based on the semantic features after association alignment and the quantization index data.

[0137] In one embodiment, the interaction scheme generation device further includes:

[0138] The compliance constraint rule acquisition module is used to acquire the compliance constraint rules in the target rule set;

[0139] The detection module is used to detect whether the interactive content in the interaction scheme is compliant based on the compliance constraint rules.

[0140] The sending module is used to send the interaction scheme to a designated terminal if the interaction content in the interaction scheme is compliant.

[0141] In one embodiment, the interaction scheme generation device further includes:

[0142] A recording module is used to record process data and result data of interaction with the specified interaction object based on the interaction scheme;

[0143] The optimization module is used to update and optimize the target rule set in the culture adaptation rule base based on the process data and the result data.

[0144] In one embodiment, the interaction scheme generation device further includes:

[0145] The original cultural feature data acquisition module is used to acquire original cultural feature data of various cultural regions;

[0146] The basic rule set generation module is used to process each of the original cultural feature data based on a preset cultural dimension quantification model, and generate a basic rule set corresponding to each cultural region.

[0147] The successful interaction case sample acquisition module is used to acquire multiple interaction scenarios and successful interaction case samples for each interaction scenario.

[0148] The successful interaction case sample analysis module is used to analyze the successful interaction case samples and extract strategy features related to the quantitative cultural dimension in the basic rule set.

[0149] The binding module is used to bind the strategy features with the corresponding cultural regions, interaction scenarios, and quantified cultural dimensions to form rule entries;

[0150] The aggregation module is used to aggregate all rule entries and build the culture-adaptive rule library.

[0151] In one embodiment, the generation module 912 includes:

[0152] The candidate interaction scheme generation submodule is used to generate multiple candidate interaction schemes based on the control conditions and the large language model.

[0153] The simulation evaluation submodule is used to perform simulation evaluation on each of the candidate interaction schemes based on a preset set of evaluation rules, and predict the expected interaction effect of each candidate interaction scheme.

[0154] The filtering submodule is used to select the target interaction scheme as the output interaction scheme from multiple candidate interaction schemes based on the expected interaction effect of each candidate interaction scheme.

[0155] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement an interaction scheme generation method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to execute the interaction scheme generation method. Those skilled in the art will understand that... Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0156] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0157] Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0158] Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base;

[0159] The target rule set is dimensionalized to obtain the interaction control parameters;

[0160] The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0161] The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model;

[0162] Based on the control conditions, an interaction scheme is generated using the large language model.

[0163] By constructing a quantifiable cultural adaptation rule base, abstract cultural dimensions are transformed into specific interaction control parameters, providing cultural constraints for the large language model and improving the cultural adaptability and regional compliance of the generated solutions. Simultaneously, by fusing and analyzing structured and unstructured business data through a multimodal data processing model, target feature vectors representing interactive objects are generated, providing integrated basis for model decision-making and enhancing the individual relevance and rationality of the solutions. Furthermore, by combining interaction control parameters and target feature vectors into structured control conditions to guide the large language model, domain knowledge is combined with model generation capabilities, achieving automation and intelligence in the interaction solution generation process, thereby improving the efficiency and consistency of strategy generation in cross-cultural business interaction scenarios.

[0164] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0165] Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data;

[0166] Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base;

[0167] The target rule set is dimensionalized to obtain the interaction control parameters;

[0168] The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object;

[0169] The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model;

[0170] Based on the control conditions, an interaction scheme is generated using the large language model.

[0171] By constructing a quantifiable cultural adaptation rule base, abstract cultural dimensions are transformed into specific interaction control parameters, providing cultural constraints for the large language model and improving the cultural adaptability and regional compliance of the generated solutions. Simultaneously, by fusing and analyzing structured and unstructured business data through a multimodal data processing model, target feature vectors representing interactive objects are generated, providing integrated basis for model decision-making and enhancing the individual relevance and rationality of the solutions. Furthermore, by combining interaction control parameters and target feature vectors into structured control conditions to guide the large language model, domain knowledge is combined with model generation capabilities, achieving automation and intelligence in the interaction solution generation process, thereby improving the efficiency and consistency of strategy generation in cross-cultural business interaction scenarios.

[0172] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0173] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0174] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for generating an interactive scheme, characterized in that, The method includes: Obtain the region identifier information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data; Based on the regional identification information, the corresponding target rule set is called from the pre-built cultural adaptation rule base; The target rule set is dimensionalized to obtain the interaction control parameters; The multi-source heterogeneous data is fused and analyzed using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object; The interactive control parameters and the target feature vector are used as control conditions and input into the pre-trained large language model; Based on the control conditions, an interaction scheme is generated using the large language model; The step of generating an interaction scheme based on the control conditions and the large language model includes: Based on the control conditions, multiple candidate interaction schemes are generated using the large language model; Based on a preset set of evaluation rules, each candidate interaction scheme is simulated and evaluated to predict the expected interaction effect of each candidate interaction scheme. Based on the expected interaction effect of each candidate interaction scheme, the target interaction scheme is selected from the multiple candidate interaction schemes as the output interaction scheme.

2. The interaction scheme generation method according to claim 1, characterized in that, The step of dimensionalizing the target rule set to obtain the interaction control parameters includes: Extract the cultural dimension theoretical indicators from the target rule set, and the weight coefficients corresponding to each cultural dimension theoretical indicator; The theoretical indicators of each cultural dimension and their corresponding weight coefficients are mapped to the interactive control parameters according to a preset mapping method.

3. The interaction scheme generation method according to claim 1, characterized in that, The step of fusing and analyzing the multi-source heterogeneous data using a preset multimodal data processing model to generate a target feature vector representing the specified interactive object includes: The unstructured business data in the multi-source heterogeneous data is parsed using a natural language processing model to obtain semantic features, and the structured business data in the multi-source heterogeneous data is extracted to obtain quantitative indicator data. The semantic features are associated and aligned with the quantitative indicator data; The target feature vector is generated based on the semantic features after association alignment and the quantitative index data.

4. The interaction scheme generation method according to claim 1, characterized in that, After the step of generating an interaction scheme based on the control conditions and the large language model, the method further includes: Obtain the compliance constraint rules from the target rule set; The compliance constraint rules are used to detect whether the interactive content in the interaction scheme is compliant. If the interactive content in the interaction scheme is compliant, the interaction scheme will be sent to the designated terminal.

5. The interaction scheme generation method according to claim 1, characterized in that, After the step of generating an interaction scheme based on the control conditions and the large language model, the method further includes: Record the process data and result data of the interaction with the specified interaction object based on the interaction scheme; Based on the process data and the result data, the target rule set corresponding to the culture adaptation rule base is updated and optimized.

6. The interaction scheme generation method according to claim 1, characterized in that, Before the step of calling the corresponding target rule set from the pre-built cultural adaptation rule base based on the regional identification information, the method further includes: Obtain original cultural characteristic data for each cultural region; Based on the preset cultural dimension quantification model, the original cultural feature data of each region are processed to generate a basic rule set corresponding to each cultural region. Acquire multiple interaction scenarios and successful interaction case samples for each scenario; Analyze the successful interaction case samples and extract strategy features related to the quantitative cultural dimension in the basic rule set; The strategy features are bound to the corresponding cultural regions, interaction scenarios, and quantified cultural dimensions to form rule entries; All rule entries are compiled to build the culture-adaptive rule base.

7. An interactive scheme generation device, characterized in that, The device includes: The acquisition module is used to acquire the region identification information of a specified interactive object, as well as multi-source heterogeneous data associated with the specified interactive object; wherein, the multi-source heterogeneous data includes structured business data and unstructured business data; The calling module is used to call the corresponding target rule set from the pre-built culture adaptation rule base based on the region identification information; The quantization module is used to quantize the dimensions of the target rule set to obtain interactive control parameters; The analysis module is used to perform fusion analysis on the multi-source heterogeneous data through a preset multimodal data processing model to generate a target feature vector representing the specified interactive object; The input module is used to input the interactive control parameters and the target feature vector as control conditions into the pre-trained large language model; The generation module is used to generate an interaction scheme based on the control conditions and the large language model; The generation module includes: The candidate interaction scheme generation submodule is used to generate multiple candidate interaction schemes based on the control conditions and the large language model. The simulation evaluation submodule is used to perform simulation evaluation on each of the candidate interaction schemes based on a preset set of evaluation rules, and predict the expected interaction effect of each candidate interaction scheme. The filtering submodule is used to select the target interaction scheme as the output interaction scheme from multiple candidate interaction schemes based on the expected interaction effect of each candidate interaction scheme.

8. A computer-readable storage medium, characterized in that, The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the interaction scheme generation method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the interaction scheme generation method as described in any one of claims 1 to 6.

Images