A micro-marketing auxiliary method and system based on large language model

Abstract

The present invention discloses a micro-enterprise marketing auxiliary method and system based on a large language model, which uses a large language model to improve the efficiency and effectiveness of marketing. The method integrates a vector database, an embedding model and contextual learning technology to achieve real-time understanding of user intentions, personalized responses and the ability to quickly adapt to business changes. The core advantages of this method lie in its accurate user demand analysis, personalized service provision and efficient knowledge retrieval system. By analyzing user behavior and conversation records, the present invention can provide product recommendations and marketing strategies that meet user preferences. In addition, the present invention also has the ability to update real-time speech to ensure the timeliness and accuracy of information; the automated order information extraction technology improves information processing efficiency and reduces the time and cost of manual operations. The present invention not only improves marketing efficiency, but also enhances the interactive experience between users and customer service, providing companies with a competitive advantage in a highly competitive market.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent marketing technology, and in particular, relates to an enterprise micro-marketing auxiliary method and system based on a large language model. Background Art

[0002] With the advent of the digital age, WeChat for Business has become an important platform for enterprises to connect with customers and promote products and services. However, although traditional marketing methods, such as offline promotion and advertising, are effective for a certain period of time, their efficiency and effectiveness are gradually insufficient in the face of growing user demand and rapidly changing market environment. Traditional marketing methods often lack accurate user positioning and personalized services, resulting in high marketing costs and low conversion rates, which are difficult to meet the pursuit of efficient and precise marketing in modern business. With the advancement of artificial intelligence technology, especially the rapid development of large language models, new opportunities have emerged in the field of intelligent marketing. These models can interact with users in a more natural and accurate way through the WeChat for Business platform, understand user intentions, and assist customer service personnel in effective communication, thereby meeting the needs of specific sales scenarios and promoting the transformation of enterprises to digitalization and intelligence.

[0003] Despite this, existing marketing assistance methods based on WeChat for Business still have some shortcomings. First, when faced with a large number of users, it is difficult for a single customer service representative to respond to all messages quickly, resulting in slow response and affecting user experience. Second, standardized answers that rely on preset script databases often fail to meet the personalized needs of users, because these methods have limitations in understanding specific business scenarios, resulting in lengthy output content that does not meet the simplicity and clarity pursued in customer service conversations. In addition, the update speed of the preset script database often lags behind business changes and the development of user problems, resulting in the information provided being outdated or inaccurate.

[0004] To overcome these challenges, a more intelligent and efficient marketing assistance method is needed that can understand user intent in real time, generate personalized responses, and quickly adapt to the changing business environment. This can not only improve marketing efficiency, but also enhance the interactive experience between users and customer service, thereby gaining an advantage in the fiercely competitive market. Summary of the invention

[0005] The present invention provides an enterprise micro-marketing assistance method and system based on a large language model, aiming to solve the problem that customer service lacks personalized answers, professional knowledge is updated slowly, user experience is reduced, and marketing efficiency is affected.

[0006] The technical solution provided by the present invention is specifically as follows:

[0007] An enterprise micro-marketing auxiliary method based on a large language model comprises the following steps:

[0008] S1. Understand business scenarios and generate personalized responses through agent-assisted methods;

[0009] S2. Build user portraits based on historical chat records to capture users’ purchasing intentions;

[0010] S3. Generate recommended words based on chat records and business scenarios to promote order progress;

[0011] S4. Identify key information fields from the chat records and generate complete user order information.

[0012] Furthermore, step S1 specifically includes:

[0013] S11, extracting chat records from the database, using the vector database to retrieve similar chat scenes, and providing them as chat examples to the large language model;

[0014] S12, the large language model generates an attack point label by analyzing and reasoning the current chat record, and the attack point is predefined in advance according to the corresponding business scenario;

[0015] S13. The large language model generates personalized responses that meet the business scenarios based on the input chat examples, attack point labels, and current chat records.

[0016] Furthermore, before retrieving similar chat scenes, the extracted chat records need to be denoised to remove irrelevant information; the retrieval of similar chat scenes requires converting the chat records into vector representations, and then using a similarity calculation method to find similar chat scenes in the vector database.

[0017] Furthermore, step S2 specifically includes:

[0018] S21. The large language model captures the user's characteristics and needs based on the chat record data and generates user recommendation tags;

[0019] S22, using the follow-up mechanism to obtain more detailed user information, obtain the corresponding label score, and use the weighted label score to evaluate the user's purchase intention, and add a new shopping intention label based on the original label;

[0020] S23. Combine the newly added tags, the recommended tags obtained from the chat record analysis, and the original user tags to intelligently generate a complete user portrait.

[0021] Furthermore, the follow-up mechanism obtains more detailed user information through a series of preset questions, and the preset questions are dynamically adjusted according to the user's answers; when intelligently generating user portraits, graph neural networks are used to capture the relationship between user tags, and graph embedding technology is used to generate low-dimensional user feature representations.

[0022] Furthermore, step S3 specifically includes:

[0023] S31. Perform fine-grained interpretation on existing business scenarios, and use the large language model to extract follow-up scenarios that match the current chat context based on chat records;

[0024] S32: The large language model integrates chat examples, attack point labels, and speech history to generate new agent-assisted speech, providing reference speech for advancing the order process.

[0025] S33. Based on the understanding of the current follow-up context and the reference response script, perform text style transfer, optimize and rewrite the generated script content to make it consistent with the tone and identity of the customer service, while maintaining the accuracy and completeness of the information.

[0026] Furthermore, after generating new agent-assisted dialogue, multi-round dialogue management technology is used to manage the dialogue process to ensure the coherence and logic of the dialogue; the multi-round dialogue management technology includes a rule-based dialogue management system or an end-to-end dialogue generation model.

[0027] Furthermore, step S4 specifically includes:

[0028] S41, scanning the entire chat record, identifying and marking text fragments that meet a specific pattern;

[0029] S42, grabbing data of a specified field from the text segment containing the required information field;

[0030] S43. Combine the captured field data, supplementary additional knowledge and a small number of explanatory examples, and use the large language model to generate complete user order information.

[0031] An enterprise micro-marketing assistance system based on the above method includes a user interaction unit, an agent assistance unit, a user portrait construction unit, a follow-up speech generation unit and an order intelligent generation unit:

[0032] The user interaction unit includes an interaction module for providing an interaction interface between the user and the customer service, and an intention recognition module for recognizing the user's intention and needs according to the user input;

[0033] The agent assistance unit includes a speech recommendation module, a knowledge retrieval module and a personalized response module: the speech recommendation module is used to recommend appropriate speech templates according to user intentions and conversation history; the knowledge retrieval module is used to quickly retrieve knowledge base content; and the personalized response module is used to generate personalized responses that meet user needs;

[0034] The user portrait construction unit includes a label generation module for generating user labels according to user behaviors and conversation records, a portrait refinement module for refining user portraits through multi-type input and questioning mechanisms, and a purchase intention evaluation module for evaluating user purchase intentions;

[0035] The follow-up speech generation unit includes a scene understanding module, a speech optimization module and a style transfer module: the scene understanding module is used to understand the current conversation scene and generate corresponding follow-up speech; the speech optimization module is used to optimize the speech content according to user feedback and conversation history; the style transfer module is used to adjust the speech style to make it consistent with the customer service identity and tone;

[0036] The order intelligent generation unit includes an information identification module, a data capture module and an order generation module: the information identification module is used to scan chat records and identify key information fields; the data capture module is used to accurately capture data from the identified fields; and the order generation module is used to generate complete user order information.

[0037] Compared with the prior art, the present invention has at least the following beneficial effects:

[0038] 1. Improve natural language processing capabilities: This invention uses the powerful natural language understanding and generation capabilities of the large language model to accurately capture user intentions and generate personalized responses that meet user needs. This not only improves the response speed of customer service, but also enhances the interactive experience between users and customer service.

[0039] 2. Achieve efficient knowledge retrieval: By integrating the vector database, the present invention can quickly retrieve the knowledge base content related to the user's query and provide accurate answers. This efficient knowledge retrieval mechanism ensures the timeliness and accuracy of information and helps to improve user satisfaction.

[0040] 3. Enhanced personalized marketing: By building a comprehensive user profile, the present invention can deeply analyze user behavior and preferences and provide personalized product recommendations and marketing strategies. This personalized marketing method not only improves conversion rates, but also enhances user loyalty to the brand.

[0041] 4. Real-time update and optimization of the script: Using contextual learning technology, the present invention can update and optimize the script content in real time according to the latest conversation content and business changes. This ensures that customer service replies always keep the latest and most relevant information, improving the effectiveness of marketing activities.

[0042] 5. Improve information processing efficiency: The present invention uses automated order information extraction technology to quickly and accurately extract key information from chat records and generate complete order data. This not only reduces the time and cost of manual operations, but also reduces the risk of errors.

[0043] 6. Strong adaptability: The method of the present invention has strong adaptability and can be adjusted and optimized according to different business scenarios and user needs. This flexibility enables the technical solution to be widely used in various marketing scenarios to meet the needs of different enterprises.

[0044] 7. Data-driven decision support: By collecting and analyzing user interaction data, the present invention provides valuable data support for enterprises, helping them make more accurate market forecasts and decisions. This data-driven decision-making method helps enterprises better grasp market dynamics and improve their competitiveness.

[0045] 8. Reduce operating costs: Through automated and intelligent marketing assistance, the present invention can significantly reduce the operating costs of enterprises in customer service and marketing. This allows enterprises to invest more resources in product development and market expansion, thereby achieving a higher return on investment.

[0046] To sum up, the present invention successfully integrates multiple technologies such as vector database, embedding model, context learning, etc., solves the problems of low efficiency, low personalization, slow response, etc. in traditional marketing, and performs well in processing natural language understanding and generation, knowledge retrieval, dialogue management, etc. It can meet the diverse needs of user consultation, order management, etc., realizes dynamic information collection, improves marketing efficiency, and the customer service replies are very close to the actual chat style, which enhances the interactive experience between users and customer service, and has certain practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0048] Figure 1 It is a schematic diagram of the overall framework of the enterprise micro-marketing assistance method provided by one embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a process of seat assistance provided by an embodiment of the present invention;

[0050] Figure 3 It is a schematic diagram of a process of constructing a user portrait provided by an embodiment of the present invention;

[0051] Figure 4 It is a schematic diagram of the process of generating follow-up speech provided by an embodiment of the present invention;

[0052] Figure 5 It is a schematic diagram of the process of extracting order information provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative work are all within the scope of protection of the present invention.

[0054] Embodiment 1

[0055] This embodiment provides a micro-enterprise marketing assistance method based on a large language model, which is mainly implemented by collecting and analyzing user chat records and prompting to call the large language model. Figure 1 As shown, it mainly includes the following steps:

[0056] (1) Agent assistance: Generate labels and scripts based on chat records to understand business scenarios and chat styles.

[0057] In the process of communication between users and customer service, the corresponding scenarios are analyzed, and corresponding chat words are provided based on the understanding of user needs in order to increase the user's transaction intention. Based on this, this implementation provides a seat assistance method, referring to Figure 2 , mainly including:

[0058] Step S11: Extract chat records from the database, use natural language processing techniques such as word segmentation, part-of-speech tagging, named entity recognition, regular expressions, etc. to remove irrelevant noise information, use the vector database to retrieve similar chat scenes, and provide them as chat examples to the large language model. The retrieval of the vector database can be done by using the BGE-M3 embedding model to convert the chat records into vector representations, and then use similarity calculation methods (such as cosine similarity) to perform sparse, dense, and ColBERT hybrid retrieval to find similar chat scenes.

[0059] It should be noted that the retrieval of similar scenes requires encoding the chat record information into a high-dimensional vector and performing a distance comparison with other encoded chat record information. Chat information in similar scenes can be captured through encoding and comparison. When removing irrelevant noise information, the Qwen2-72B-Instruct model can also be used for context understanding to more accurately identify and filter out information that is irrelevant to the current conversation. The retrieval of the vector database can also be combined with semantic similarity search technology, such as using Milvus and Elasticsearch efficient similarity search libraries to speed up the retrieval process.

[0060] Step S12: The large language model generates attack point labels based on the chat record context to provide business knowledge for the model.

[0061] Attack points can be predefined in advance according to the corresponding business scenarios. The large language model integrates and infers the analysis of chat records and the provided attack point information to obtain the corresponding attack point information. When the large model generates attack point labels, it can use its ability to understand the context and combine the information in the business knowledge base to infer the products or services that the user may be interested in. For example, if a user asks for information about a certain product, the model can recognize that this is a sales opportunity and generate corresponding labels. When generating attack point labels, multi-task learning technology can be used to allow the model to learn multiple related tasks (such as intent recognition and sentiment analysis) at the same time to improve the accuracy and relevance of label generation. In addition, knowledge graphs can be introduced to enhance the model's understanding of business knowledge.

[0062] Step S13: Fusion chat examples, attack point labels, and speech history to generate new chat speech, and provide auxiliary functions that meet business scenarios. When fusing chat examples, attack point labels, and speech history, the Qwen2-72B-Instruct model can be used to generate new conversation speech. The model can learn the best response methods in different scenarios, and combine the user's historical conversation records to generate personalized responses. After generating new conversation speech, post-processing techniques such as text summarization and keyword extraction can be used to ensure the conciseness and emphasis of the speech.

[0063] After adopting agent-assisted technology, the model's responses to similar business scenarios are shorter, it understands the user's personalized needs, has basic user response skills, and is more in line with the usual customer service chat style.

[0064] (2) Build user profiles, utilize label classification of multiple types of inputs, and use a questioning mechanism to refine user information to capture the user’s purchasing intention.

[0065] In the marketing process, in order to help companies more accurately grasp user needs, improve product user experience, optimize marketing strategies, and thus increase the attractiveness of products and services, it is necessary to build user portraits based on relevant information. Based on this, this implementation provides a method for building user portraits. Figure 3 , mainly including:

[0066] Step S21: The large language model generates appropriate user tags based on the processed chat record data, captures user characteristics and urgent needs, and then derives the corresponding offensive point words. When generating user tags, the recursive structure construction tree, compressed node content and semantic association search can be used to analyze the user's conversation content and identify the user's interests, purchasing habits and other characteristics. In addition, external data sources (such as social media data) can be combined to enrich user portraits.

[0067] Step S22: Use a label classification method that includes multiple types of input (such as selection boxes and text input boxes) to refine and quantify user label features. Use a follow-up questioning mechanism to dig deeper into the details of the user portrait. Use offensive point words to ask questions to obtain more detailed user information and obtain corresponding label scores. Sum label scores with different weights to evaluate the user's willingness to buy. Add some labels with high shopping intentions on the basis of the original labels. The follow-up questioning mechanism can dynamically adjust questions according to the user's answers by designing a series of preset questions to obtain more detailed information.

[0068] Step S23: Combine the newly added partial tags, the recommended tags obtained from the chat record analysis, and the original user tags to intelligently generate a complete user portrait, ensure the real-time update of the portrait, and realize a comprehensive, dynamic, and intelligent user portrait construction service. When intelligently generating user portraits, Graph Neural Networks can be used to capture the relationship between user tags, and Graph Embedding can be used to generate low-dimensional user feature representations. In addition, reinforcement learning algorithms can be introduced to allow the model to continuously learn in the interaction with users and optimize the user portrait construction process.

[0069] The user portrait technology is used to capture the user's immediate characteristics and urgent needs, provide detailed user characteristic analysis for generating targeted attack point words, realize dynamic information collection, and also improve the level of personalization of marketing and services.

[0070] (3) Follow up on the generation of scripts, generate recommended scripts based on chat records, reference scripts and chat scenarios, and promote the order process.

[0071] In the process of communicating with users, in order to promote the order process more efficiently, customer service needs to generate appropriate follow-up words according to the current chat context. Based on this, this implementation provides a method for generating follow-up words. Figure 4 , which mainly include:

[0072] Step S31: Perform fine-grained interpretation of existing business scenarios. The large language model extracts follow-up scenarios based on chat records to ensure that they are consistent with the current context and facilitate the subsequent provision of business scenario knowledge. When interpreting business scenarios in a fine-grained manner, knowledge graph technology can be used to structure business knowledge and combine it with the Qwen2-72B-Instruct model to generate explanatory text. In addition, multimodal learning technology can be used to combine non-text data such as images and audio to enhance the model's understanding of business scenarios.

[0073] Step S32: The large language model integrates chat examples, attack point labels and speech history to generate new agent-assisted speech, which provides reference speech for advancing the order process. After generating new agent-assisted speech, multi-round dialogue management technology, such as rule-based dialogue management system or end-to-end dialogue generation model, can be used to manage the dialogue process and ensure the coherence and logic of the speech.

[0074] Step S33: Based on the understanding of the current follow-up context and the reference speech reply, perform text style transfer, optimize and rewrite the generated speech content to make it consistent with the tone and identity of the customer service, while maintaining the accuracy and completeness of the information. When performing text style transfer, models such as generative adversarial networks (GAN) or variational autoencoders (VAE) can be used to allow the model to learn text features of different styles and generate speech content that conforms to a specific style. In addition, the speech generation model can be continuously adjusted and optimized in combination with user feedback and evaluation.

[0075] By adopting the follow-up script generation technology, we focus on business scenarios and improve the style. The customer service replies are very close to the actual chat style, while retaining the retrieved business knowledge and improving the overall marketing efficiency.

[0076] (4) Order information extraction: identify key information fields from chat records based on a small number of examples and accurately extract comprehensive order information.

[0077] When processing a large number of user chat records, in order to quickly and accurately obtain order information, it is necessary to design an efficient order information extraction method. Based on this, this implementation provides an order information extraction method. Figure 5 , the main steps include:

[0078] Step S41: Scan the entire chat record, identify and mark the text fragments that meet the specific pattern, so that the subsequent steps can quickly locate and extract relevant information. When scanning the chat record, the Qwen2-72B-Instruct model and Elasticsearch full-text search can be used to identify and mark key information fields. In addition, specific business rules, order stages, and speech keywords can be combined to improve the accuracy of recognition.

[0079] Step S42: Accurately capture the data of the specified field from the sentence containing the required information field. For information such as time, the field data is calculated and formatted according to the conversation date. For information such as location and appearance, directly extract relevant words or phrases, and the last one in the chat record shall prevail.

[0080] Step S43: Combine the extracted field data, supplementary additional knowledge and a small number of illustrative examples, use the large language model to generate complete user order information, and format it into a JSON object to facilitate subsequent data processing or analysis.

[0081] The use of order information extraction technology effectively avoids the need for manual browsing of lengthy chat records to manually extract order information in traditional marketing methods, greatly improves the response efficiency and accuracy of information processing, and reduces the possibility of human error.

[0082] The above is the main content of the enterprise micro-marketing auxiliary method provided by this embodiment. To verify its effect, this embodiment is deployed and implemented based on the Qwen2 large language model in the experimental environment of Ubuntu 20.04.6 LTS operating system, Python 3.10 programming environment, AMD EPYC 7742 64-Core CPU, 512G memory, and 8-core RTX 3090 GPU. The results show that it can meet the diverse needs of user consultation, order management, etc., realize dynamic information collection, provide personalized answers that meet specific scenarios, enhance the interactive experience between users and customer service, and improve marketing efficiency.

[0083] Embodiment 2

[0084] Based on the above method, this embodiment provides a micro-enterprise marketing assistance system based on a large language model, which aims to improve marketing efficiency and effectiveness through large language model technology. The following is a detailed description of the functional units that the system should include and its workflow.

[0085] 1. System functional units

[0086] 1. User interaction unit

[0087] Interaction module: provides an interactive interface between users and customer service, supporting multiple forms of communication such as text, voice, and pictures.

[0088] Intent recognition module: Analyze user input in real time and identify user intentions and needs.

[0089] 2. Agent Assistance Unit

[0090] Speech recommendation module: recommends appropriate speech templates based on user intent and conversation history.

[0091] Knowledge retrieval module: quickly retrieve relevant product information, frequently asked questions and other knowledge base content.

[0092] Personalized response module: Generates personalized responses that meet user needs.

[0093] 3. User portrait construction unit

[0094] Label generation module: Generate user labels based on user behavior and conversation records.

[0095] Portrait refinement module: refines user portraits through multi-type input and questioning mechanisms.

[0096] Purchase intention evaluation module: evaluates user purchase intention and provides a basis for marketing strategies.

[0097] 4. Follow-up script generation unit

[0098] Scenario understanding module: understands the current conversation scenario and generates corresponding follow-up words.

[0099] Speech optimization module: optimizes speech content based on user feedback and conversation history.

[0100] Style transfer module: adjust the speaking style to make it consistent with the customer service identity and tone.

[0101] 5. Order intelligent generation unit

[0102] Information identification module: scans chat records and identifies key information fields.

[0103] Data capture module: accurately capture data from identified fields.

[0104] Order generation module: Generates complete user order information and formats it as a JSON object.

[0105] 2. System Workflow

[0106] When a user asks a question about a product, the user interaction module first captures the question and passes it to the intent recognition component. After analyzing the intent of the question, the intent recognition component passes the result to the agent assistance module. Based on the intent, the agent assistance module retrieves relevant information from the knowledge base and generates a personalized response. At the same time, the user portrait construction module is also analyzing the conversation content, generating user tags, and evaluating the purchase intention. This information is passed to the agent assistance module and the follow-up speech generation module to generate responses and follow-up speech that better meet user needs. When the user expresses the intention to purchase, the order information extraction module starts working, extracts the order information, and generates complete order data.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may also be combined, the steps may be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A micro-enterprise marketing assistance method based on a large language model, characterized in that: Includes steps: S1. Understand business scenarios and generate personalized responses through agent-assisted methods: S11, extracting chat records from the database, using the vector database to retrieve similar chat scenes, and providing them as chat examples to the large language model; S12, the large language model generates an attack point label by analyzing and reasoning the current chat record, and the attack point is predefined in advance according to the corresponding business scenario; S13: The large language model generates personalized responses that meet the business scenario based on the input chat examples, attack point labels, and current chat records; S2. Build user portraits based on historical chat records to capture users’ purchasing intentions: S21. The large language model captures the user's characteristics and needs based on the chat record data and generates user recommendation tags; S22, using the follow-up mechanism to obtain more detailed user information, obtain the corresponding label score, and use the weighted label score to evaluate the user's purchase intention, and add a new shopping intention label based on the original label; S23, combining the newly added tags, the recommended tags obtained by analyzing the chat records, and the original user tags to intelligently generate a complete user profile; S3. Generate recommended words based on chat records and business scenarios to promote order progress; S4. Identify key information fields from the chat records and generate complete user order information.

2. The micro-enterprise marketing assistance method according to claim 1, characterized in that: Before retrieving similar chat scenes, the extracted chat records need to be denoised to remove irrelevant information; Retrieval of similar chat scenes requires converting chat records into vector representations and then using similarity calculation methods to find similar chat scenes in the vector database.

3. The micro-enterprise marketing assistance method according to claim 1, characterized in that: The follow-up mechanism obtains more detailed user information through a series of preset questions, and the preset questions are dynamically adjusted according to the user's answers; when intelligently generating user portraits, a graph neural network is used to capture the relationship between user tags, and graph embedding technology is used to generate a low-dimensional user feature representation.

4. The micro-enterprise marketing assistance method according to claim 1, characterized in that: Step S3 specifically includes: S31. Perform fine-grained interpretation on existing business scenarios, and use the large language model to extract follow-up scenarios that match the current chat context based on chat records; S32: The large language model integrates chat examples, attack point labels, and speech history to generate new agent-assisted speech, providing reference speech for advancing the order process. S33. Based on the understanding of the current follow-up context and the reference response script, perform text style transfer, optimize and rewrite the generated script content to make it consistent with the tone and identity of the customer service, while maintaining the accuracy and completeness of the information.

5. The enterprise micro-marketing auxiliary method according to claim 4, characterized in that: After generating new agent-assisted dialogue, multi-round dialogue management technology is used to manage the dialogue process to ensure the coherence and logic of the dialogue; the multi-round dialogue management technology includes a rule-based dialogue management system or an end-to-end dialogue generation model.

6. The micro-enterprise marketing assistance method according to claim 1, characterized in that: Step S4 specifically includes: S41, scanning the entire chat record, identifying and marking text fragments that meet a specific pattern; S42, grabbing data of a specified field from the text segment containing the required information field; S43. Combine the captured field data, supplementary additional knowledge and a small number of explanatory examples, and use the large language model to generate complete user order information.

7. An enterprise micro-marketing assistance system based on the method according to any one of claims 1 to 6, characterized in that: It includes user interaction unit, agent assistance unit, user portrait construction unit, follow-up speech generation unit and order intelligent generation unit; The user interaction unit includes an interaction module for providing an interaction interface between the user and the customer service, and an intention recognition module for recognizing the user's intention and needs according to the user input; The agent assistance unit includes a speech recommendation module, a knowledge retrieval module and a personalized response module: the speech recommendation module is used to recommend appropriate speech templates according to user intentions and conversation history; the knowledge retrieval module is used to quickly retrieve knowledge base content; and the personalized response module is used to generate personalized responses that meet user needs; The user portrait construction unit includes a label generation module for generating user labels according to user behaviors and conversation records, a portrait refinement module for refining user portraits through multi-type input and questioning mechanisms, and a purchase intention evaluation module for evaluating user purchase intentions; The follow-up speech generation unit includes a scene understanding module, a speech optimization module and a style transfer module: the scene understanding module is used to understand the current conversation scene and generate corresponding follow-up speech; the speech optimization module is used to optimize the speech content according to user feedback and conversation history; the style transfer module is used to adjust the speech style to make it consistent with the customer service identity and tone; The order intelligent generation unit includes an information identification module, a data capture module and an order generation module: the information identification module is used to scan chat records and identify key information fields; the data capture module is used to accurately capture data from the identified fields; and the order generation module is used to generate complete user order information.

Images