A Multi-Dimensional Context Fusion AI Instruction Generation Method and System for Foreign Trade Business

By using a multi-dimensional contextual fusion AI instruction generation method, the issues of accuracy and efficiency in content generation in foreign trade AI communication tools have been resolved, enabling accurate matching and efficient generation of content in foreign trade B2B business.

CN123068549APending Publication Date: 2026-11-10HIGH HOPE AGLORY LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611241925.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing AI communication tools for foreign trade lack the ability to understand multi-dimensional business contexts, resulting in generated content that cannot accurately match customer characteristics, communication history, and product knowledge, thus failing to meet the requirements of foreign trade B2B business for accuracy and efficiency in communication.

Method used

By receiving natural language commands, the system acquires scenario setting parameters, customer profile data, historical communication style data, and product knowledge data, performs multi-stage fusion, constructs comprehensive prompt words, calls a large language model to generate content, and combines business rule verification and manual review for confirmation.

Benefits of technology

It improves the consistency between generated content and customer business needs and profile characteristics, reduces the frequency of manual intervention in integrating multi-source background information, and ensures that generated content complies with business rules and is consistent with historical high-quality communication style.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068549A_ABST
    Figure CN123068549A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-dimensional contextual fusion AI instruction generation method and system for foreign trade business, involving the fields of artificial intelligence natural language processing and foreign trade business informatization. The method includes: receiving natural language instructions input by a user; obtaining current business scenario setting parameters; querying customer profile data of the target customer; retrieving historical communication style data; obtaining product knowledge data; fusing the natural language instructions, scenario setting parameters, customer profile data, historical communication style data, and product knowledge data in multiple stages during runtime to construct a comprehensive prompt word according to a structured format; sending the comprehensive prompt word to a large language model, calling the large language model to execute content generation; performing business rule verification on the generated result, and presenting the verified generated result to the user for review and confirmation. This invention improves the consistency between the generated content and customer business needs and profile characteristics, reducing the frequency of manual intervention in integrating multi-source background information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence natural language processing and foreign trade business informatization. Specifically, it relates to a multi-dimensional context fusion AI instruction generation method and system for foreign trade business. Background Technology

[0002] With the digitalization of global trade, artificial intelligence (AI) dialogue systems are increasingly being applied to communication support in B2B foreign trade. However, B2B foreign trade communication has significant industry-specific characteristics: on the one hand, customers are distributed across different countries and regions, with significant differences in their procurement stages, cultural backgrounds, and communication habits; on the other hand, a single business communication often requires comprehensive consideration of heterogeneous information from multiple sources, such as customers' historical transaction records, current order status, and product specifications.

[0003] Existing general-purpose AI dialogue systems are primarily trained on general corpora and lack a deep understanding of the business context within the foreign trade vertical. In practical applications, the responses generated by such systems are often overly generalized and fail to accurately match the characteristics and needs of specific customers. To obtain usable responses, business personnel typically need to manually retrieve and organize contextual data such as customer background and product information, then concatenate it into prompts for input into the AI ​​system. This process is cumbersome and time-consuming, severely hindering the efficiency of business communication.

[0004] Furthermore, most foreign trade support tools currently on the market use fixed templates or simple slot-filling techniques to generate emails and messages. These tools cannot adaptively adjust based on multi-dimensional information such as specific customer profiles, historical communication styles, and dynamically updated product knowledge, resulting in rigid and untargeted content that fails to meet the actual needs of personalized and refined communication in foreign trade B2B business.

[0005] In summary, existing technologies lack an effective mechanism for automatically integrating multi-dimensional business context information in generating communication content for foreign trade business. This makes it difficult for AI-generated content to accurately match customer characteristics, communication history, and product knowledge, thus failing to meet the requirements of foreign trade B2B business for accuracy and efficiency in communication. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the shortcomings of existing foreign trade AI communication tools, such as weak contextual understanding and lack of targeted content generation. It aims to solve the problem of how to automatically integrate multi-dimensional business contextual information to accurately match customer characteristics, communication history and product knowledge when AI generates foreign trade business communication content. The invention proposes a multi-dimensional contextual fusion AI instruction generation method and system for foreign trade business.

[0007] To address the aforementioned technical problems, according to one aspect of the present invention, a multi-dimensional context fusion AI instruction generation method for foreign trade business is provided, the method comprising:

[0008] S1. Receive natural language instructions input by the user, wherein the natural language instructions contain a description of the user's business requirements for AI-generated content;

[0009] S2. Obtain the current business scenario setting parameters, which include the identity settings, tone style and dialogue strategy of multiple preset business scenario types;

[0010] S3. Query the customer profile data of the target customer, which includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics.

[0011] S4. Retrieve historical communication style data, which includes target quality response samples associated with the target customer and their corresponding weight values;

[0012] S5. Obtain product knowledge data, including product specifications, pricing information, and certifications.

[0013] S6. Natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data are integrated in multiple stages during runtime, and comprehensive prompt words are constructed based on the integration results in a structured format.

[0014] S7. Send the comprehensive prompt words to the large language model, call the large language model to perform content generation, and output the generated results;

[0015] S8. Perform business rule verification on the generated results, and present the verified generated results to the user for review and confirmation, so as to realize the generation of AI instructions with multi-dimensional context fusion.

[0016] According to another aspect of the present invention, a multi-dimensional context fusion AI instruction generation system for foreign trade business is also provided, the system comprising:

[0017] The natural language instruction receiving module is used to receive natural language instructions input by the user, wherein the natural language instructions contain the user's business requirements description of the AI-generated content;

[0018] The scenario setting parameter acquisition module is used to acquire the current business scenario setting parameters, which include the identity settings, tone style and dialogue strategy of multiple preset business scenario types.

[0019] The customer profile data query module is used to query the customer profile data of target customers. The customer profile data includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics.

[0020] The historical communication style data retrieval module is used to retrieve historical communication style data, which includes target quality response samples associated with target customers and their corresponding weight values.

[0021] The product knowledge data acquisition module is used to acquire product knowledge data, which includes product specifications, pricing information, and certification qualifications.

[0022] The data fusion and prompt word construction module is used to integrate natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data in multiple stages during runtime, and construct comprehensive prompt words based on the fusion results in a structured format.

[0023] The large language model invocation module is used to send comprehensive prompt words to the large language model, invoke the large language model to perform content generation, and output the generated results;

[0024] The verification and result presentation module is used to perform business rule verification on the generated results and present the verified results to the user for review and confirmation, so as to realize the generation of AI instructions through multi-dimensional context fusion.

[0025] The beneficial effects of this invention are as follows:

[0026] 1. This invention integrates natural language instructions, scene setting parameters, customer profile data, historical communication style data, and product knowledge data in multiple stages during runtime to construct a business context association graph and perform conflict resolution and dynamic weight allocation. This enables comprehensive prompts to simultaneously carry business needs, identity constraints, object characteristics, style references, and content knowledge, thereby improving the consistency between generated content and customer business needs and profile characteristics, and reducing the frequency of manual intervention in integrating multi-source background information.

[0027] 2. This invention uses a pre-trained quality scoring model to score candidate reply samples in historical emails in multiple dimensions, selects target quality reply samples and stores them in a style sample library and assigns weight values. In the prompt word construction stage, historical communication style data is selected as input for the style reference layer based on the weight values, thereby improving the consistency between the generated content and the historical high-quality communication style.

[0028] 3. This invention establishes a final manual review mechanism for generated content by performing business rule verification after content generation and presenting the verified generation results to the user for review and confirmation, ensuring that the output content complies with business rules and is used only after manual authorization. Attached Figure Description

[0029] The invention will now be further described with reference to the accompanying drawings.

[0030] Figure 1 This is one of the overall flowcharts of an example of the present invention;

[0031] Figure 2 This is the second overall flowchart of an example of the present invention;

[0032] Figure 3 This is a principle block diagram of an example of the present invention.

[0033] In the picture:

[0034] 1. Natural Language Command Receiving Module; 2. Scene Setting Parameter Acquisition Module; 3. Customer Profile Data Query Module; 4. Historical Communication Style Data Retrieval Module; 5. Product Knowledge Data Acquisition Module; 6. Data Fusion and Prompt Word Construction Module; 7. Large Language Model Calling Module; 8. Validation and Result Presentation Module. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings, which will make the technical approach and operation steps of the present invention clearer.

[0036] According to embodiments of the present invention, a method and system for generating multi-dimensional context fusion AI instructions for foreign trade business are provided.

[0037] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1-2 As shown, according to an embodiment of the present invention, a multi-dimensional context fusion AI instruction generation method for foreign trade business is provided, the method comprising:

[0038] S1. Receive natural language instructions input by the user, which contain the user's business requirements for AI-generated content;

[0039] It should be added that users can input natural language commands through the AI ​​(artificial intelligence) input box in the email composition interface, such as "urging for samples, emphasizing the tight delivery schedule," and the system will use this command as core command layer data.

[0040] S2. Obtain the current business scenario setting parameters. The scenario setting parameters include the identity settings, tone style and dialogue strategy of multiple preset business scenario types.

[0041] It should be added that the parameters of the currently selected scenario are obtained from 12 pre-set business scenarios; the scenarios are divided into four groups: email processing group, including replying to customers, order follow-up, sample follow-up, etc.; customer development group, including development emails, re-engaging dormant customers, holiday greetings, etc.; customer analysis group, including customer profile analysis, communication strategy suggestions, etc.; product group, including product recommendation, quotation follow-up, new product promotion, etc.; each scenario has pre-set identity settings, tone style, and communication strategy parameters.

[0042] S3. Query the target customer's customer profile data. The customer profile data includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics.

[0043] It should be added that the structured profile data of the target customers can be retrieved from the preset customer database, including their country and region of origin; current procurement stage, such as inquiry, sampling, trial order, and repeat order; historical communication style preferences, such as formal, concise, and friendly; and decision-making speed characteristics, such as fast, cautious, and process-oriented.

[0044] S4. Retrieve historical communication style data, which includes target quality response samples associated with the target customer and their corresponding weight values.

[0045] In this optional embodiment, S4 includes:

[0046] S41. Based on the target customer identifier, retrieve historical email records with the target customer from the preset email database, and based on the sender's identity identifier, filter out reply emails sent by business personnel from the historical email records to generate a candidate reply sample set.

[0047] S42. Based on the candidate response sample set, calculate the sub-scores of each candidate response sample using the pre-trained quality scoring model, and sum them according to the preset dimension weight coefficients to obtain the comprehensive score; among which, the sub-scores include semantic relevance score, response completeness score, style consistency score, business rule compliance score, and language quality score.

[0048] In this optional embodiment, S42 includes:

[0049] S421. For each candidate response sample in the candidate response sample set, perform text cleaning, word segmentation and vectorization encoding operations in sequence to generate response text vector;

[0050] S422. Extract customer emails corresponding to candidate reply samples from historical email records, and perform text cleaning, word segmentation and vectorization encoding operations on the customer emails in sequence to generate email text vectors.

[0051] S423. Concatenate and fuse the reply text vector with the incoming email text vector to construct paired input features, and input the paired input features into the pre-trained quality scoring model.

[0052] S424. Utilize the output heads of multiple evaluation dimensions in the quality scoring model to perform forward inference calculations, and output semantic relevance score, response completeness score, style consistency score, business rule compliance score, and language quality score respectively, generating a set of sub-scores;

[0053] S425. The scores of each sub-item in the sub-score set are weighted and summed according to the preset dimension weight coefficients to obtain the comprehensive score of each candidate response sample.

[0054] S43. Mark the candidate response samples with a comprehensive score higher than the preset score threshold as target quality response samples, store them in the preset style sample library, and assign weight values ​​to the target quality response samples in the style sample library according to the preset mapping rules, so as to establish an association index between target customers and target quality response samples.

[0055] S44. Based on the target customer identifier, retrieve the associated target quality response samples from the style sample library, sort them in descending order of weight value, and select the top preset number of target quality response samples as historical communication style data.

[0056] It should be added that historical high-quality response samples (i.e., target quality response samples) relevant to the target customer are retrieved from a pre-defined style sample library. The style sample library stores high-quality responses selected from historical emails using a quality scoring model, with each sample accompanied by a weight value. The search results are sorted by weight, and the top N samples are selected as style references.

[0057] The quality scoring model's evaluation dimensions include: semantic relevance (whether the response content matches the customer's business needs in the email, weighted at 30%); response completeness (whether it covers all the questions and needs raised by the customer, weighted at 25%); style consistency (whether the tone and expression of the response match the customer's communication style, weighted at 20%); business rule compliance (whether the professional content such as price, delivery time, and product specifications in the response is accurate, weighted at 15%); and language quality (whether the grammar, spelling, and format are correct, weighted at 10%). The overall score for each email is the weighted sum of the scores from each dimension. Emails with a score higher than 80 (out of 100) are marked as high-quality samples, stored in the style sample library, and assigned corresponding weight values. The scoring model can be continuously optimized based on user feedback. Retrieving historical communication style data: The system retrieves historical high-quality response samples related to the target customer from the style sample library. The style sample library stores high-quality responses selected from historical emails using the quality scoring model, with each sample accompanied by a weight value. The search results are sorted by weight, and the top N samples are selected as style references. Furthermore, the quality scoring model is built based on a BERT pre-trained model. Each reply was labeled with a score of 1-5 on over 60,000 historical foreign trade emails. The model then employs the Adam optimizer and a cross-entropy loss function (learning rate 2×10⁻⁶). −5 batch size Supervised training was performed with 32 epochs and 5 epochs, and the F1 score on the test set reached 0.89.

[0058] In one specific embodiment, all email records with historical correspondence with the target customer are retrieved from a pre-set email database based on the target customer identifier. This email database stores over 100,000 foreign trade emails, each containing a sender identification field. Reply emails sent by business personnel are selected based on the identification, generating a candidate reply sample set containing 80 emails. A quality score is performed on each candidate reply sample in the candidate reply sample set. Specifically, the candidate reply samples are sequentially cleaned to remove signatures and special symbols; segmented using BERT's Word Piece segmenter; and the segmentation results are input into a pre-trained BERT-base model to extract the 768-dimensional vector corresponding to the [CLS] tag as the reply text vector. Simultaneously, the customer's incoming emails that triggered the reply are extracted from historical email records, and similarly cleaned, segmented, and encoded using BERT to generate a 768-dimensional incoming email text vector. The reply text vector and the incoming email text vector are concatenated and fused to construct a 1536-dimensional paired input feature. Paired input features are fed into a pre-trained quality scoring model, which is built on a BERT pre-trained model and trained under supervision on over 60,000 historical foreign trade emails. The training data consists of scores ranging from 1 to 5, assigned to each reply by experienced business personnel across five dimensions: semantic relevance, reply completeness, style consistency, compliance with business rules, and language quality. Training employs the Adam optimizer and cross-entropy loss function, with a learning rate of 2×10⁻⁶. −5With a batch size of 32, after training for 5 epochs, the model achieved an F1 score of 0.89 on the test set. The model contains five independent evaluation dimension output heads, each of which is a fully connected classification layer that maps 1536-dimensional features to 5 class probabilities. The expected value of the probability distribution is used to obtain the original scores of dimensions 1 to 5, which are then multiplied by 20 and linearly converted into percentage-based sub-scores. This simultaneously outputs semantic relevance score, response completeness score, style consistency score, business rule compliance score, and language quality score, forming a set of sub-scores. The comprehensive score of a candidate response sample is obtained by weighting and summing the scores according to the preset dimensional weight coefficients: semantic relevance (30%), response completeness (25%), style consistency (20%), business rule compliance (15%), and language quality (10%). For example, if the scores of a response in the five dimensions are 4, 5, 4, 5, and 4, they are converted to 80, 100, 80, 100, and 80, respectively. The weighted calculation is 0.3×80 + 0.25×100 + 0.2×80 + 0.15×100 + 0.1×80 = 88 points. After scoring all candidate response samples, samples with a comprehensive score higher than the preset scoring threshold of 80 points are marked as target quality response samples, totaling 22. These are stored in a preset style sample library, and according to the preset mapping rules, the comprehensive score is directly normalized to the [0,1] interval as a weight value. For example, 88 points corresponds to a weight of 0.88. A weight value is assigned to each target quality response sample, and an association index is established between the target customer identifier and these samples and their weights. When historical communication style data needs to be retrieved, all associated target quality response samples are retrieved from the style sample library based on the current target customer identifier. These samples are then sorted in descending order of weight value, and the top 5 samples are selected as the final output historical communication style data. For example, the weights of the selected 5 samples are 0.95, 0.92, 0.88, 0.85, and 0.80, respectively. These samples and their weight values ​​will be directly carried into the subsequent fusion process for the calculation of the information specificity factor, ensuring that the high-quality feedback in the historical communication style data can effectively influence the generation weight of the prompt words.

[0059] S5. Obtain product knowledge data, including product specifications, pricing information, and certifications.

[0060] It is necessary to add that structured knowledge data of the target product should be obtained from the product database, including product name, specifications, price range, minimum order quantity, certifications, etc.

[0061] S6. Natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data are integrated in multiple stages during runtime, and comprehensive prompt words are constructed based on the integration results in a structured format.

[0062] In this optional embodiment, S6 includes:

[0063] S61. Standardize and clean natural language commands, scenario setting parameters, customer profile data, historical communication style data, and product knowledge data, and extract target information items.

[0064] S62. Perform named entity recognition on the target information item, extract business entities, and use business entities as anchors to perform cross-data source layer association mapping to construct a business context association graph containing entity nodes and relationship edges.

[0065] S63. Based on the business context association graph, traverse and generate candidate conflict items, and perform conflict intensity quantification and classification resolution on the candidate conflict items. Based on the resolution results, perform resolution confidence verification and output conflict-free information items.

[0066] In this optional embodiment, S63 includes:

[0067] S631. Traverse the nodes in the business context association graph, extract target information items with the same anchor point but different source data level, and generate candidate conflict items.

[0068] S632. Based on the differences in attribute values ​​of candidate conflict items, calculate the conflict intensity score, and based on the preset intensity range in which the conflict intensity score is located, execute the corresponding hierarchical resolution strategy to generate resolution information items.

[0069] In this optional embodiment, S632 includes:

[0070] S6321. Extract attribute values ​​from candidate conflict items, and calculate the semantic distance and numerical deviation between different attribute values ​​to obtain the conflict intensity score.

[0071] S6322. Compare the conflict intensity score with the preset conflict threshold. Based on the comparison result, mark the candidate conflict items with conflict intensity scores higher than the preset conflict threshold as substantive conflict items, and mark the candidate conflict items with conflict intensity scores not higher than the preset conflict threshold as non-substantive conflict items.

[0072] S6323. Compare the conflict intensity score of the substantive conflict item with the preset intensity range to determine the target intensity range in which the substantive conflict item is located; wherein, the intensity range includes a first intensity range, a second intensity range and a third intensity range;

[0073] S6324. Based on the determined target intensity range, match the corresponding hierarchical resolution strategy, perform resolution processing on the substantive conflict items, and output the valid information items.

[0074] In this optional embodiment, S6324 includes:

[0075] If the target intensity range is the first intensity range, the dynamic data source priority sequence under the current scenario is determined according to the preset scenario priority adjustment factor, and the attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is selected from the substantive conflict items as the valid information item of the substantive conflict item.

[0076] If the target intensity range is the second intensity range, the dynamic data source priority sequence under the current scenario is determined according to the preset scenario priority adjustment factor. The attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is selected from the substantive conflict items as the basic valid value. The attribute value with the second highest priority is selected from the remaining substantive conflict items as the candidate valid value. The basic valid value and the candidate valid value are used together as the valid information item of the substantive conflict item.

[0077] If the target intensity range is the third intensity range, then the substantive conflict item is marked as an abnormal conflict item, an abnormal pending confirmation mark is attached, and a pending confirmation task is generated and submitted to the user. Simultaneously, the attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is used as a temporary placeholder valid value. After the user confirms, the user confirmation value replaces the temporary placeholder valid value as the valid information item of the substantive conflict item.

[0078] S6325. Merge non-substantive conflict items with valid information items to generate resolution information items.

[0079] S633. Calculate the resolution confidence level based on the resolution information item and business context association graph, compare the resolution confidence level with the preset confidence threshold, and generate conflict-free information items based on the comparison results.

[0080] In this optional embodiment, S633 includes:

[0081] S6331. Use the business context association graph to extract the association path of the business entities associated with the information item to be resolved, and calculate the semantic consistency score between the attribute value of the information item to be resolved and the attribute value of the associated node on the association path, as the resolution confidence.

[0082] S6332. Compare the resolution confidence level with the preset confidence level threshold to determine whether the resolution confidence level is lower than the preset confidence level threshold. If so, mark the resolution information item as an information item to be confirmed and add a conflict doubt mark to the information item to be confirmed. If not, treat the resolution information item as a confirmed conflict-free item.

[0083] S6333: Summarize the information items to be confirmed, the items without conflict, and the non-substantial conflict items, and generate the information items without conflict.

[0084] S64. Based on the conflict resolution state of the conflict-free information items, determine the corresponding state adjustment coefficient, calculate the semantic importance score based on the state adjustment coefficient, and generate candidate information items with weight labels by combining the preset score weight mapping function.

[0085] In this optional embodiment, S64 includes:

[0086] S641. Analyze whether there is a conflict doubt flag for the conflict-free information item, mark the conflict resolution status of the conflict-free information item based on the analysis results, and determine the corresponding status adjustment coefficient; wherein, the conflict resolution status includes doubt status and confirmed status.

[0087] S642. Calculate semantic importance sub-factors based on the association graph between conflict-free information items and business context; among which, semantic importance sub-factors include entity relevance factor, information specificity factor, and timeliness factor.

[0088] S643. The semantic importance sub-factors are weighted and fused according to the preset sub-factor weight coefficients to generate an initial semantic importance score, and the initial semantic importance score is dynamically adjusted in combination with the state adjustment coefficient to obtain the semantic importance score.

[0089] S644. Based on the preset score weight mapping function, the semantic importance score is converted into a dynamic fusion weight, and the dynamic fusion weight is associated with the conflict-free information item to generate a candidate information item with a weight label.

[0090] S65. Using natural language instructions as the target instruction layer, scenario setting parameters as the identity constraint layer, customer profile data as the object feature layer, historical communication style data as the style reference layer, and product knowledge data as the content knowledge layer, candidate information items are sequentially arranged according to a preset hierarchical order, and the arrangement result is embedded into a preset structured prompt word template to generate comprehensive prompt words.

[0091] It should be added that the five layers of data are processed according to a multi-stage fusion process. The first stage is hierarchical data preprocessing, which standardizes and cleans the data at each layer, extracting key information fields. The second stage is semantic alignment, which maps the five layers of data to business entities (such as customer name, product SKU, and order number) to ensure that each layer points to the same business context. The third stage is semantic importance scoring, which scores the information items in each layer based on factors including business entity relevance, information specificity, and historical interaction frequency. The fourth stage is weighted fusion, which weights and superimposes the importance scores of each layer of data with preset layer weights to form a unified intermediate representation. The fifth stage is conflict detection and resolution, which detects contradictions between different layers of data (such as price parameter conflicts) and automatically selects or marks them for processing according to priority rules. The sixth stage is structured formatting, which formats the fused data according to a structured Prompt template to generate the final comprehensive Prompt.

[0092] The scoring factors include: Entity relevance, which determines whether the information item contains key business entities such as customer name, product SKU, and price (i.e., entity relevance factor); Information specificity, which measures the uniqueness and irreplaceability of the information item in the current context (i.e., information specificity factor); and Historical utility, which measures the contribution of this type of information item to historical high-quality responses (i.e., timeliness factor). The information specificity factor, for information items derived from historical communication style data, is further weighted and adjusted by incorporating the weight values ​​of the target quality response samples associated with that information item. ,in Basic specificity, These are the sample weight values; the three factors correspond to the weight coefficients of the weighted fusion components. , , .

[0093] The semantic importance score is used to adjust the weight allocation of data at each layer during fusion, with high-importance information items receiving higher fusion weights. Simultaneously, low-importance information items (semantic importance scores below a preset threshold) are automatically pruned to reduce redundant information in the fused prompt, improving the accuracy and efficiency of the large language model's generation. Result review and confirmation: The system returns the output generated by the large language model to the email composition interface for business personnel to review and modify. The system can only be used after business personnel confirmation, reflecting the human-led, AI-assisted approach. The semantic importance score is calculated using the formula... Calculation, where For the first The semantic importance score of each information item, with a value range of [value range missing]. , Entity relevance score (calculated based on named entity recognition technology). The semantic relevance score is calculated based on semantic similarity. This is a timeliness factor (the more recent, the higher). Weighting coefficients (default) ).

[0094] The specific method for constructing the business context association graph is as follows: Business entities extracted by named entity recognition are used as entity nodes in the graph, including customer entities, product entities, order entities, communication style entities, etc. Target information items corresponding to the same entity in different data sources are used as attributes of that entity node, and the attribute source layer is labeled with the data source type. When different entities have a co-occurrence relationship in the same data source, such as mentioning both product X and price Y in an email, a relationship edge is established between the corresponding entity nodes. The type of the relationship edge is labeled as a co-occurrence relationship, and the weight of the edge is determined by the co-occurrence frequency after TF-IDF normalization. The final result is a business context association graph containing multiple entity nodes, attribute key-value pairs, and weighted relationship edges. Its data structure is stored in the form of an adjacency list, and each node maintains a list of attribute sources.

[0095] The association path refers to the set of paths formed by performing a breadth-first traversal along the relationship edges in the business context association graph, starting from the business entity associated with the information item being resolved, to obtain all adjacent entity nodes and their attribute values ​​with a path depth not exceeding 2 hops. When calculating the resolution confidence, the attribute value text of all adjacent nodes on the association path is extracted, and cosine similarity is calculated between each attribute value text and the attribute value text of the information item being resolved. The weighted average of all similarities is taken as the semantic consistency score, with the weight being the reciprocal of the path length from each adjacent node to the starting entity node. For example, if the attribute value "value-oriented" of the information item being resolved has a similarity of 0.78 with the 1-hop node "data-supported quotation" and a similarity of 0.65 with the 2-hop node "direct and rigorous," then the semantic consistency score = (0.78 × 1 + 0.65 × 0.5) / (1 + 0.5) = 0.737, which is used as the resolution confidence.

[0096] This implementation adds a prompt word optimization step after the fusion processing and before sending it to the large language model. The semantic information entropy of each word in the current Prompt is calculated. ,in For a given word Lexical elements in context The system uses the conditional probability distribution to identify redundant and key semantic regions in the Prompt. Simultaneously, it extracts the distribution features from the attention heatmap. The system identifies key regions where attention is highly concentrated. Based on a comprehensive analysis of semantic information entropy and self-attention heatmap distribution, the system performs dynamic lexical pruning. The system also incorporates adversarial vector anchoring technology. Constraining the generation boundary in the latent space effectively prevents attention dilution and style assimilation during the autoregressive generation process of long texts.

[0097] In one specific embodiment, standardized cleaning is performed. This involves processing the following parameters: Natural language instructions (please respond to a customer in a certain country's inquiry about product X's delivery time and price in a professional and friendly tone); scenario settings (identity as a sales representative with a professional and friendly tone; value-oriented communication strategy); customer profile data showing the target customer is located in a certain country, in the business negotiation stage, with a direct and rigorous communication style and a moderate decision-making speed; historical communication style data consisting of five target quality response samples retrieved from a style sample library with weights of 0.95, 0.92, 0.88, 0.85, and 0.80); and product knowledge data indicating product X as an industrial controller with a unit price of $100. The data includes CE certification and a standard delivery time of 30 days. Target information items were extracted from each sample. Natural language commands yielded four intent items: "Reply," "Product X," "Delivery Time," and "Price." Scenario setting parameters yielded "Identity: Sales Representative," "Tone: Professional and Friendly," and "Strategy: Value-Oriented." Customer profiles yielded "Country: [Country Name]," "Procurement Stage: Negotiation," "Style Preference: Direct and Rigorous," and "Decision Speed: Medium." Historical communication style data extracted frequently occurring sentence patterns from five samples, such as data comparison and factual statements. Product knowledge was extracted, including "Product X," "Price: $100," "Delivery Time: 30 Days," and "Certification: CE." All information items were labeled with their source data layer. Named entity recognition was performed on the target information items to extract business entities such as Product X, Customer A, Price, and Delivery Time. Using these entities as anchors, related attributes across five layers of data were mapped to construct a business context association graph. In the graph, the entity node Product X is associated with "Price: $100" and "Delivery Time: 30 days" from the Product Knowledge layer. Multiple responses from the Historical Style layer consistently emphasized the implicit attribute of delivery time stability. The "medium decision-making speed" from the customer profile is associated with entity Customer A. The "Strategy: Value-Oriented" from the Scenario Setting layer and the factual data priority pattern extracted from the Historical Style layer are both associated with the virtual node "Communication Style," forming cross-layer connections. Based on this graph, candidate conflict items were generated: It was found that under the Communication Style node, the value-oriented attribute value from the Scenario Setting layer and the attribute value extracted from five samples from the Historical Style layer exhibited potential contradictions. These were identified as candidate conflict items with the same anchor but different source data layers. For this conflict item, the vector representations corresponding to the two attribute values ​​are extracted. The value orientation and data comparison statement are encoded into 768-dimensional vectors using a pre-trained BERT model. The cosine similarity is calculated to be 0.62, and the semantic distance is 1-0.62=0.38. Since the two are text descriptions with no direct numerical deviation, the conflict intensity score is taken as the semantic distance of 0.38. The preset conflict threshold is 0.2. 0.38 is higher than the threshold and is marked as a substantial conflict item.Comparing 0.38 with preset intensity ranges: the first intensity range ([0.2, 0.4)) corresponds to mild conflict, the second intensity range ([0.4, 0.7)) corresponds to moderate conflict, and the third intensity range ([0.7, 1.0]) corresponds to severe conflict. 0.38 falls into the first intensity range, therefore, a graded resolution strategy based on the first intensity range is adopted. At this point, the dynamic data source priority sequence for the current scenario is determined based on a preset scenario priority adjustment factor. This factor is based on the current customer profile's procurement stage being business negotiation and the scenario's communication strategy being value-oriented. Through a pre-calibrated adjustment matrix, the priority weight of the scenario setting data source is increased from the default 0.2 to 0.4, the weight of the historical communication style data source is dynamically adjusted from 0.25 to 0.3, the weight of the product knowledge layer remains at 0.2, the weight of the customer profile layer is adjusted to 0.05, and the weight of the natural language instruction layer is adjusted to 0.05. The dynamic data source priority sequence is obtained in descending order of weight: Scenario Setting > Historical Communication Style > Product Knowledge > Customer Profile > Natural Language Instruction. According to this dynamic sequence, the highest priority data source, namely the attribute value "value orientation" corresponding to the scenario setting layer, is selected as the valid information item for the substantial conflict items. Meanwhile, other information items, such as the price of product X, are all $100 and conflict-free across all layers; these are directly marked as non-substantial conflict items. The valid information item "value orientation" obtained from the first intensity interval is merged with all non-substantial conflict items to generate a set of resolved information items. Confidence verification is performed on the resolved information items: the business context association graph is used to extract the association path of the business entity customer A associated with the resolved information item "value orientation". The graph shows that a price negotiation node appeared in customer A's historical emails, and the historical reply samples under this node all used a data-supported pricing communication method, which has a semantic association with "value orientation". The semantic consistency score between the attribute value "value orientation" of the resolved information item and the data-supported pricing of adjacent nodes on the association path is calculated. After text encoding, the cosine similarity is calculated to be 0.78, which is used as the resolution confidence. The preset confidence threshold is 0.7. Since 0.78 is higher than the threshold, this resolved information item is marked as a confirmed conflict-free item, without the need for an additional conflict-questioning label. All resolved information items and non-substantial conflict items are aggregated into a set of conflict-free information items, including information such as a confirmed conflict-free value-oriented strategy, a conflict-free price of $100, a delivery time of 30 days, and a direct and rigorous customer style preference.

[0098] Calculate the semantic importance score: For conflict-free information items, analyze their value orientation and identify any conflict-related doubts. The conflict resolution status is confirmed, and the corresponding status adjustment coefficient is 1.0. Calculate the semantic importance sub-factors: The entity relevance factor E assesses the direct correlation between the information item and the core business entity, customer A, and communication style. Based on graph connectivity, it is calculated to be 0.9. The information specificity factor R calculates the uniqueness of the information item in the current context, i.e., its average semantic similarity with other information items. The less similar, the higher the specificity. It is calculated to be 0.85. The timeliness factor F is the normalized value of the most recent update time of the data source to which the information item belongs. The scenario is set to be configured on the same day, so it is set to 1.0. According to the formula... Where α=0.4, β=0.4, and γ=0.2, the initial semantic importance score is calculated as 0.4×0.9+0.4×0.85+0.2×1.0=0.90, which is then multiplied by a state adjustment coefficient of 1.0 to obtain the final semantic importance score of 0.90. For other conflict-free information items such as "Price: $100", the entity relevance E=1.0, information specificity R=0.6, and timeliness F=1.0, the score is 0.4×1.0+0.4×0.6+0.2×1.0=0.84, with a state adjustment coefficient of 1.0, resulting in a final score of 0.84; "Delivery time: 30 days" is calculated similarly to obtain 0.80. Based on the preset score weight mapping function, the score is directly used as the dynamic fusion weight, which is the association weight value for each information item. For example, the value-oriented strategy weight is 0.90, the price information weight is 0.84, and the delivery time information weight is 0.80. After weighting, the system is arranged in a preset hierarchical order, with natural language instructions as the target instruction layer, scenario settings as the identity constraint layer, customer profiles as the object feature layer, historical communication styles as the style reference layer, and product knowledge as the content knowledge layer. Weighted information items are sequentially arranged, with high-weight items placed first in each layer. The arrangement results are embedded into a structured prompt template to generate comprehensive prompts in the following form: [Instruction] Generate reply email; [Identity] Sales representative, professional and friendly tone; [Customer Characteristics] Customer from a certain country, in the negotiation stage, prefers directness and rigor; [Style Reference] Prioritize value-oriented strategies (weight 0.90), supplemented by data support; [Content Knowledge] Product X, price $100 (weight 0.84), delivery time 30 days (weight 0.80), CE certification.

[0099] Before sending the comprehensive prompt word to the large language model, the system performs a prompt word optimization step: calculating the semantic information entropy of each word element in the current prompt word. The conditional probabilities are provided by the pre-trained language model. It identifies words with a professional / friendly tone and those that can be supplemented with certain information as having an information entropy higher than 0.85, indicating information redundancy. Conversely, words with an information entropy lower than 0.3, such as "product X," "$100," and "30 days," are identified as key semantic regions. Simultaneously, it extracts information from the attention heatmap distribution. The study found that the attention weight between value orientation and direct rigor reached 0.7, proving that the two are closely related and constitute core constraints, while the attention weight of redundant region lexical units averaged below 0.1. Based on comprehensive analysis, dynamic lexical unit pruning was performed, removing lexical units with information entropy higher than 0.8 and average attention weight lower than 0.15 from the prompts. After pruning, the prompts were shortened by 15% while retaining key constraints. An adversarial vector anchoring technique was introduced, adding perturbations to the hidden layer representation of each Transformer layer during the generation stage. The constraint ∥δ∥2≤0.01 makes... To strengthen the boundary constraints on professional, friendly, and value-oriented styles during the generation process, and to prevent attention dilution during autoregression of long texts from causing the generated content to converge into generic polite phrases. Optimized comprehensive prompts are sent to the large language model, and the generated reply email automatically includes a $100 quote for product X and a 30-day delivery time. The entire email adopts a customer-value-centric argumentative structure, and concludes with data listing the product's advantages while maintaining a professional and friendly tone, aligning with the intended integration after conflict resolution and weight adjustment.

[0100] S7. Send the comprehensive prompt words to the large language model, call the large language model to perform content generation, and output the generated results;

[0101] It should be added that the comprehensive prompt is sent to the Large Language Model API (Application Programming Interface). During the call, a token bucket rate limiting algorithm is used to control the request frequency, and an exponential backoff retry strategy is used to handle temporary errors. The specific parameters of the exponential backoff retry strategy are set as follows: initial retry interval 0.5 seconds, maximum number of retries 5, and backoff factor 1.5, meaning the waiting time for the nth retry is 0.5 × (1.5). n−1 Seconds; if 5 consecutive retries fail, an error message is returned to the email composing interface to prompt the user for manual handling. The large language model used in this implementation is the DeepSeek model, employing a Transformer Decoder-only architecture based on MoE technology, with a total parameter size of approximately 671B and a context window length of 128K tokens. When calling, set temperature=0.7, max... tokens =1024、top p =0.95. An adaptive capacity adjustment mechanism is introduced based on the token bucket rate limiting algorithm: when the average response time exceeds 2.0 seconds, the token replenishment rate is reduced from the default 10 times / second to 5 times / second; it is gradually restored during idle periods.

[0102] S8. Perform business rule verification on the generated results, and present the verified generated results to the user for review and confirmation, so as to realize the generation of AI instructions with multi-dimensional context fusion.

[0103] It should be added that the output generated by the large language model is returned to the email composition interface for business personnel to review and modify. The output can only be used after business personnel confirmation, reflecting the concept of human-led AI assistance. Furthermore, the system records user editing operations on the generated results, using the modified content as feedback data to update the weight values ​​of the style sample library. The specific update mechanism is as follows: the final reply email modified by the user is compared with the original generated result in terms of textual differences. If the modification exceeds a preset threshold (e.g., the edit distance exceeds 15%), the modified reply email is re-input into the S42 quality scoring model as a new candidate sample for scoring. If the score is higher than 80, it is stored in the style sample library and assigned an initial weight. For the original target quality reply sample referenced during generation, if its weight value is higher than 0.8 and the user has made significant modifications, the sample weight value is multiplied by a decay factor of 0.9 for weight reduction adjustment; if the user only makes minor modifications or uses it directly, the sample weight value is multiplied by an enhancement factor of 1.05 for weight increase adjustment, with a weight cap of 1.0.

[0104] In one specific embodiment, the business rule verification includes multiple automatic check rules based on product knowledge data and business compliance requirements. For example, the price must match the selling price of product X (US$100) in the product knowledge base with a deviation of no more than 1%; the delivery period must be accurate to 30 days and the wording must not contain ambiguous terms; certification information must include CE and must not be omitted; and a sensitive word library filters any statements containing promises or absolutes that violate regulations. The currently generated reply email text is: "Dear customer, thank you for your inquiry; Product X is a best-selling industrial controller, priced at US$100, with a typical delivery period of approximately 35 days. It has CE certification, and we believe it will bring excellent value to your production line." The verification engine scans each rule. The price of US$100 matches the rule, and the mention of CE certification is also approved. However, the delivery period rule detects that "approximately 35 days" does not match the standard delivery period of 30 days, and "approximately" is ambiguous, triggering a delivery period deviation alarm. This generated result is marked as verification failed. The system highlights "approximately 35 days" in red in the email composition interface and pops up a message indicating a delivery period deviation; it should be 30 days, and the use of vague qualifying words is prohibited. The generated result was not sent directly but presented to the business personnel for review and confirmation. Upon seeing the system prompt, the business personnel manually changed the period from approximately 35 days to 30 days and adjusted the ending tone, changing "We believe this will bring outstanding value to your production line" to "Based on the efficiency improvement indicators you previously focused on, we can provide detailed data comparisons to support your selection," making the response more in line with the client's direct and rigorous communication style and historical preference for factual data. After the user confirmed the changes and submitted, the system recorded the user's editing operations and compared the original generated text with the final modified response text. The Levenshtein edit distance algorithm was used to calculate the difference. The original text length was 238 characters, and the modified text length was 245 characters, resulting in an edit distance of 38. The edit distance ratio (38 / 245 ≈ 15.5%) exceeded the preset modification threshold of 15%, triggering the feedback update mechanism. The system uses the user's revised final reply email as a new candidate sample and re-enters it into the S42 quality scoring model for scoring. The model scores the revised reply with 4 points for semantic relevance (converted to 80 points), 5 points for reply completeness (100 points), 5 points for style consistency (100 points), 5 points for business rule compliance (100 points), and 5 points for language quality (100 points). The result is 88 points, which is higher than 80 points, calculated by weighting the dimensions with 0.3, 0.25, 0.2, 0.15, and 0.1 respectively. Therefore, the revised reply email is stored in the style sample library and given an initial weight of 0.88. At the same time, an association index with the current customer identifier is established.Meanwhile, the weights of the original five target quality response samples referenced in this generation were dynamically adjusted based on the user's modification level: the sample with a weight of 0.95, due to significant user modifications, had its weight multiplied by a decay factor of 0.9, resulting in a weight of 0.95 × 0.9 = 0.855; the sample with a weight of 0.92 was similarly multiplied by 0.9, resulting in a weight of 0.828; and the samples with weights of 0.88, 0.85, and 0.80, due to minor user modifications or low reference levels, were multiplied by an enhancement factor of 1.05, resulting in weights of 0.88 × 1.05 = 0.924 (upper limit 1.0, not reached), 0.85 × 1.05 = 0.8925 ≈ 0.89, and 0.80 × 1.05 = 0.84, respectively. In the updated style sample library, the weight distribution of samples associated with this customer more accurately reflects the characteristics of efficient responses in the latest business communications. Subsequent retrievals of historical communication style data will prioritize samples with higher weights, forming a closed loop of continuous optimization. The final reply email, which is verified and confirmed by the user, is sent by the system. Throughout the process, human and machine collaboration is maintained, and business personnel always retain the final decision-making power, achieving high-quality business communication under the AI ​​instruction generation of multi-dimensional context fusion.

[0105] To verify the technical effectiveness of this invention, a practical deployment test was conducted within a company. The system has completed the full parsing and import of over 60,000 historical emails, automatically constructing over 2,000 customer profiles covering multiple export categories such as apparel, fabrics, and home textiles. The system is pre-configured with 12 foreign trade business scenarios. In the test environment, the average time for single command fusion was 15 milliseconds, the average time for large language model requests was 1.2 seconds (including network transmission), and the adoption rate of generated replies reached 65%. The above actual operating data verifies the effectiveness and practicality of the technical solution of this invention. Result review and confirmation: The system returns the output generated by the large language model to the email composition interface, presenting it to business personnel for review and modification. The system can only be used after confirmation by business personnel, reflecting the concept of human-led AI assistance.

[0106] Comparative testing conditions: 1000 real foreign trade business scenarios were selected, and 5 senior business personnel generated responses using both the present invention and traditional template methods, respectively, and then conducted double-blind scoring. Comparative results: The present invention scored an average of 4.15 points, while the traditional template scored 2.83 points (an improvement of 46.6%); the present invention took 2.3 seconds to generate responses, while the traditional template took 5.8 seconds (an improvement of 60.3%); the present invention had a direct delivery rate of 72.1%, while the traditional template had a rate of 18.5% (an improvement of 53.6%); as shown in Table 1.

[0107] Table 1. Comparison of test results between the present invention and traditional template methods.

[0108] Evaluation indicators This invention Traditional template method Increase Double-blind scoring (average score) 4.15 points 2.83 points 46.6% Generation time 2.3 seconds 5.8 seconds 60.3% Straight hair rate 72.1% 18.5% 53.6%

[0109] like Figure 3As shown, according to another embodiment of the present invention, a multi-dimensional context fusion AI instruction generation system for foreign trade business is also provided, the system comprising:

[0110] Natural language instruction receiving module 1 is used to receive natural language instructions input by the user. The natural language instructions contain the user's business requirements for AI-generated content.

[0111] The scenario setting parameter acquisition module 2 is used to acquire the current business scenario setting parameters, which include the identity settings, tone style and dialogue strategy of multiple preset business scenario types.

[0112] Customer profile data query module 3 is used to query the customer profile data of target customers. The customer profile data includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics.

[0113] The historical communication style data retrieval module 4 is used to retrieve historical communication style data, which includes target quality response samples associated with target customers and their corresponding weight values.

[0114] Product knowledge data acquisition module 5 is used to acquire product knowledge data, which includes product specifications, pricing information and certification qualifications.

[0115] The data fusion and prompt word construction module 6 is used to integrate natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data in multiple stages during runtime, and construct comprehensive prompt words based on the fusion results in a structured format.

[0116] The large language model calling module 7 is used to send the comprehensive prompt words to the large language model, call the large language model to perform content generation, and output the generated results.

[0117] The verification and result presentation module 8 is used to perform business rule verification on the generated results and present the verified generated results to the user for review and confirmation, so as to realize the generation of AI instructions through multi-dimensional context fusion.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-dimensional context fusion AI instruction generation method for foreign trade business, characterized in that, The method includes: S1. Receive natural language instructions input by the user, wherein the natural language instructions contain a description of the user's business requirements for AI-generated content; S2. Obtain the current business scenario setting parameters, which include the identity settings, tone style and dialogue strategy of multiple preset business scenario types; S3. Query the customer profile data of the target customer, which includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics. S4. Retrieve historical communication style data, which includes target quality response samples associated with the target customer and their corresponding weight values; S5. Obtain product knowledge data, including product specifications, pricing information, and certifications. S6. Natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data are integrated in multiple stages during runtime, and comprehensive prompt words are constructed based on the integration results in a structured format. S7. Send the comprehensive prompt words to the large language model, call the large language model to perform content generation, and output the generated results; S8. Perform business rule verification on the generated results, and present the verified generated results to the user for review and confirmation, so as to realize the generation of AI instructions with multi-dimensional context fusion.

2. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 1, characterized in that, S4 includes: S41. Based on the target customer identifier, retrieve historical email records with the target customer from the preset email database, and based on the sender's identity identifier, filter out reply emails sent by business personnel from the historical email records to generate a candidate reply sample set. S42. Based on the candidate response sample set, calculate the sub-scores of each candidate response sample using a pre-trained quality scoring model, and sum them according to the preset dimension weight coefficients to obtain a comprehensive score; wherein, the sub-scores include semantic relevance score, response completeness score, style consistency score, business rule compliance score, and language quality score. S43. Mark the candidate response samples with a comprehensive score higher than the preset score threshold as target quality response samples, store them in the preset style sample library, and assign weight values ​​to the target quality response samples in the style sample library according to the preset mapping rules, so as to establish an association index between target customers and target quality response samples. S44. Based on the target customer identifier, retrieve the associated target quality response samples from the style sample library, sort them in descending order of weight value, and select the top preset number of target quality response samples as historical communication style data.

3. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 2, characterized in that, S42 includes: S421. For each candidate response sample in the candidate response sample set, perform text cleaning, word segmentation and vectorization encoding operations in sequence to generate response text vector; S422. Extract customer emails corresponding to candidate reply samples from historical email records, and perform text cleaning, word segmentation and vectorization encoding operations on the customer emails in sequence to generate email text vectors. S423. Concatenate and fuse the reply text vector with the incoming email text vector to construct paired input features, and input the paired input features into the pre-trained quality scoring model. S424. Utilize the output heads of multiple evaluation dimensions in the quality scoring model to perform forward inference calculations, and output semantic relevance score, response completeness score, style consistency score, business rule compliance score, and language quality score respectively, generating a set of sub-scores; S425. The scores of each sub-item in the sub-score set are weighted and summed according to the preset dimension weight coefficients to obtain the comprehensive score of each candidate response sample.

4. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 1, characterized in that, S6 includes: S61. Standardize and clean natural language commands, scenario setting parameters, customer profile data, historical communication style data, and product knowledge data, and extract target information items. S62. Perform named entity recognition on the target information item, extract business entities, and use business entities as anchors to perform cross-data source layer association mapping to construct a business context association graph containing entity nodes and relationship edges. S63. Based on the business context association graph, traverse and generate candidate conflict items, and perform conflict intensity quantification and classification resolution on the candidate conflict items. Based on the resolution results, perform resolution confidence verification and output conflict-free information items. S64. Based on the conflict resolution state of the conflict-free information items, determine the corresponding state adjustment coefficient, calculate the semantic importance score based on the state adjustment coefficient, and generate candidate information items with weight labels by combining the preset score weight mapping function. S65. Using natural language instructions as the target instruction layer, scenario setting parameters as the identity constraint layer, customer profile data as the object feature layer, historical communication style data as the style reference layer, and product knowledge data as the content knowledge layer, candidate information items are sequentially arranged according to a preset hierarchical order, and the arrangement result is embedded into a preset structured prompt word template to generate comprehensive prompt words.

5. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 4, characterized in that, S63 includes: S631. Traverse the nodes in the business context association graph, extract target information items with the same anchor point but different source data level, and generate candidate conflict items. S632. Based on the differences in attribute values ​​of candidate conflict items, calculate the conflict intensity score, and based on the preset intensity range in which the conflict intensity score is located, execute the corresponding hierarchical resolution strategy to generate resolution information items. S633. Calculate the resolution confidence level based on the resolution information item and business context association graph, compare the resolution confidence level with the preset confidence threshold, and generate conflict-free information items based on the comparison results.

6. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 5, characterized in that, The S632 includes: S6321. Extract attribute values ​​from candidate conflict items, and calculate the semantic distance and numerical deviation between different attribute values ​​to obtain the conflict intensity score. S6322. Compare the conflict intensity score with the preset conflict threshold. Based on the comparison result, mark the candidate conflict items with conflict intensity scores higher than the preset conflict threshold as substantive conflict items, and mark the candidate conflict items with conflict intensity scores not higher than the preset conflict threshold as non-substantive conflict items. S6323. Compare the conflict intensity score of the substantive conflict item with a preset intensity range to determine the target intensity range in which the substantive conflict item is located; wherein, the intensity range includes a first intensity range, a second intensity range and a third intensity range; S6324. Based on the determined target intensity range, match the corresponding hierarchical resolution strategy, perform resolution processing on the substantive conflict items, and output the valid information items. S6325. Merge non-substantive conflict items with valid information items to generate resolution information items.

7. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 6, characterized in that, The S6324 includes: If the target intensity range is the first intensity range, the dynamic data source priority sequence under the current scenario is determined according to the preset scenario priority adjustment factor, and the attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is selected from the substantive conflict items as the valid information item of the substantive conflict item. If the target intensity range is the second intensity range, the dynamic data source priority sequence under the current scenario is determined according to the preset scenario priority adjustment factor. The attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is selected from the substantive conflict items as the basic valid value. The attribute value with the second highest priority is selected from the remaining substantive conflict items as the candidate valid value. The basic valid value and the candidate valid value are used together as the valid information item of the substantive conflict item. If the target intensity range is the third intensity range, then the substantive conflict item is marked as an abnormal conflict item, an abnormal pending confirmation mark is attached, and a pending confirmation task is generated and submitted to the user. Simultaneously, the attribute value corresponding to the highest priority data source in the dynamic data source priority sequence is used as a temporary placeholder valid value. After the user confirms, the user confirmation value replaces the temporary placeholder valid value as the valid information item of the substantive conflict item.

8. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 5, characterized in that, The S633 includes: S6331. Use the business context association graph to extract the association path of the business entities associated with the information item to be resolved, and calculate the semantic consistency score between the attribute value of the information item to be resolved and the attribute value of the associated node on the association path, as the resolution confidence. S6332. Compare the resolution confidence level with the preset confidence level threshold to determine whether the resolution confidence level is lower than the preset confidence level threshold. If so, mark the resolution information item as an information item to be confirmed and add a conflict doubt mark to the information item to be confirmed. If not, treat the resolution information item as a confirmed conflict-free item. S6333: Summarize the information items to be confirmed, the items without conflict, and the non-substantial conflict items, and generate the information items without conflict.

9. The multi-dimensional context fusion AI instruction generation method for foreign trade business according to claim 4, characterized in that, S64 includes: S641. Analyze whether there is a conflict doubt flag for the conflict-free information item, mark the conflict resolution status for the conflict-free information item based on the analysis results, and determine the corresponding status adjustment coefficient; wherein, the conflict resolution status includes doubt status and confirmation status. S642. Calculate semantic importance sub-factors based on the association graph between conflict-free information items and business context; wherein, the semantic importance sub-factors include entity relevance factors, information specificity factors, and timeliness factors. S643. The semantic importance sub-factors are weighted and fused according to the preset sub-factor weight coefficients to generate an initial semantic importance score, and the initial semantic importance score is dynamically adjusted in combination with the state adjustment coefficient to obtain the semantic importance score. S644. Based on the preset score weight mapping function, the semantic importance score is converted into a dynamic fusion weight, and the dynamic fusion weight is associated with the conflict-free information item to generate a candidate information item with a weight label.

10. A multi-dimensional context fusion AI instruction generation system for foreign trade business, used to implement the multi-dimensional context fusion AI instruction generation method for foreign trade business as described in any one of claims 1-9, characterized in that, The system includes: The natural language instruction receiving module is used to receive natural language instructions input by the user, wherein the natural language instructions contain the user's business requirements description of the AI-generated content; The scenario setting parameter acquisition module is used to acquire the current business scenario setting parameters, which include the identity settings, tone style and dialogue strategy of multiple preset business scenario types. The customer profile data query module is used to query the customer profile data of target customers. The customer profile data includes the target customer's country / region, purchasing stage, communication style preferences, and decision-making speed characteristics. The historical communication style data retrieval module is used to retrieve historical communication style data, which includes target quality response samples associated with target customers and their corresponding weight values. The product knowledge data acquisition module is used to acquire product knowledge data, which includes product specifications, pricing information, and certification qualifications. The data fusion and prompt word construction module is used to integrate natural language commands, scenario setting parameters, customer profile data, historical communication style data and product knowledge data in multiple stages during runtime, and construct comprehensive prompt words based on the fusion results in a structured format. The large language model invocation module is used to send comprehensive prompt words to the large language model, invoke the large language model to perform content generation, and output the generated results; The verification and result presentation module is used to perform business rule verification on the generated results and present the verified results to the user for review and confirmation, so as to realize the generation of AI instructions through multi-dimensional context fusion.