Intelligent customer service response method and device, equipment and medium

Abstract

The invention relates to an intelligent customer service response method and device, equipment and a medium, and the method comprises the steps: responding to a user question event, and obtaining a latest short-time chat record, the user question event comprising a latest question statement; according to the short-time chat records, determining historical behavior data and long-time chat records matched with the semantics of the user; combining the short-time chat record, the long-time chat record and the historical behavior data into question reference data, and determining an intention commodity and intention intensity thereof according to the question reference data; and the question reference data, the intention commodity and the intention intensity of the intention commodity are embedded into a response prompt text, response information is generated according to response service logic in the response prompt text, and the response information comprises own commodities determined from the online shop according to the intention commodity. According to the method, the accuracy and the individuation level of the response information are improved, the flexibility and the effect of commodity recommendation are enhanced, and remarkable commercial value and user satisfaction improvement are brought to online shops of an e-commerce platform.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to an intelligent customer service response method and its device, equipment, and medium. Background Art

[0002] E-commerce platforms have introduced intelligent customer service response systems for online stores, which can greatly improve the efficiency and quality of customer service. Most of the earlier intelligent customer service response systems relied on preset question-answer pairs or simple keyword matching to generate responses, and they were obviously insufficient in handling complex conversations and understanding the user's deep intentions. The real needs of users are often hidden behind their questions, and the customer service system needs to be able to deeply understand the user's semantics and behavior patterns in order to provide more accurate and personalized services.

[0003] Currently, intelligent customer service systems have developed to the stage of using user portrait tags and / or chat record tags to generate response information. E-commerce platforms can use this to predict users' intention to purchase products and recommend products based on user intentions. Among them, user portraits collect multi-dimensional information data of users, and chat record tags are essentially for generating user portrait tags, such as demographic attributes, social attributes, behavioral preferences, and consumption habits, to abstract the overall picture of users as the basis for recommending products.

[0004] In practice, although user portraits provide a way of personalized service, the product recommendation business of e-commerce platforms has higher requirements for recommending products in response information. Therefore, the intelligent customer service system that generates response information based on user portraits still has some limitations when applied to the product recommendation scenarios of e-commerce platform online stores.

[0005] First, for product recommendations in online stores, the existing intelligent customer service systems are relatively mechanical and one-sided in their recommendation logic. These systems often simply match products with semantically similar semantics to the user's profile for recommendation, while ignoring the user's real needs and purchase motivations, or the user's intended product intention strength is not clear enough to make responsive processing in the response information. This recommendation method is likely to cause users to be indifferent and cannot effectively improve users' willingness to buy.

[0006] Secondly, the user portrait labels extracted based on behavioral data and chat records are often too abstract and often cannot provide comprehensive semantic information, resulting in the matched response information not accurately corresponding to the user's true intentions, thus affecting service quality and user satisfaction.

[0007] Furthermore, the background reference information composed of user portrait tags and chat record tags that the existing technology relies on when generating response information is not comprehensive enough, and lacks intelligent decision-making logic when finally forming the response information. This leads to insufficient flexibility when recommending products, poor recommendation results, and cannot effectively improve the sales targets of online stores. Summary of the invention

[0008] The purpose of this application is to solve the above-mentioned problems and to provide an intelligent customer service response method and its corresponding devices, equipment, non-volatile readable storage medium, and computer program product.

[0009] According to one aspect of the present application, an intelligent customer service response method is provided, comprising:

[0010] In response to a user question event, the latest short chat record between the chat robot and the user is obtained, wherein the user question event includes the latest question statement of the user;

[0011] According to the short chat records, determining semantically compatible historical behavior data from the user's behavior data, and determining semantically compatible long chat records from the user's historical chat records;

[0012] The short chat records, long chat records and historical behavior data are combined into question reference data, and the intended product that the user intends to consult and its corresponding intention strength are determined according to the question reference data;

[0013] The question reference data, intended products and their intention strength are embedded in a preset answer prompt text, and according to the preset answer business logic of the answer prompt text, answer information corresponding to the intention strength and matching the latest question statement is generated, and the answer information includes the owned products determined from the online store based on the intended products.

[0014] According to another aspect of the present application, there is provided an intelligent customer service answering device, comprising:

[0015] A question response module is configured to respond to a user question event and obtain the latest short chat record between the chat robot and the user, wherein the user question event includes the latest question statement of the user;

[0016] A data matching module, configured to determine semantically matching historical behavior data from the user's behavior data based on the short-term chat records, and to determine semantically matching long-term chat records from the user's historical chat records;

[0017] an intention identification module, configured to combine the short chat records, long chat records and historical behavior data into question reference data, and determine the intended product that the user intends to consult and its corresponding intention strength according to the question reference data;

[0018] The response generation module is configured to embed the question reference data, intended products and their intention strength into a preset response prompt text, and generate response information corresponding to the intention strength and matching the latest question statement according to the preset response business logic of the response prompt text, wherein the response information includes the own products determined from the online store based on the intended products.

[0019] According to another aspect of the present application, an intelligent customer service answering device is provided, comprising a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the method described in the present application.

[0020] According to another aspect of the present application, a non-volatile readable storage medium is provided, which stores a computer program implemented according to the intelligent customer service response method in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the method are executed.

[0021] According to another aspect of the present application, a computer program product is provided, comprising a computer program / instruction, wherein the computer program / instruction implements the steps of the method when executed by a processor.

[0022] Compared with the prior art, this application has the following beneficial effects:

[0023] First, this application can more accurately capture the user's true intentions and needs by comprehensively analyzing the user's short chat records, historical behavior data, and long chat records. This multi-dimensional data fusion method effectively overcomes the problem that the user portrait labels in the prior art are too abstract and cannot provide comprehensive semantic information, so that the generated response information can more accurately correspond to the user's true intentions, improving service quality and user satisfaction.

[0024] Secondly, this application is more sophisticated and flexible in terms of product recommendation logic. It not only takes into account the user's behavioral data and the semantic similarity between the chat records and the question statements, but also comprehensively considers the short-term chat records and long-term chat records in the human-computer dialogue process. It also further clarifies the user's intended products and their intention strength, and makes corresponding recommendations in the response information accordingly, providing different recommendation schemes for different intention strengths. This recommendation method based on intention strength can better stimulate users' willingness to buy and effectively improve the sales conversion rate of online stores.

[0025] In addition, the application guides the generation of response information through the response business logic in the preset response prompt text, making the generation process of response information more intelligent. The application of this intelligent decision-making logic significantly improves the flexibility of recommending products and improves the recommendation results, thereby more effectively promoting the sales goals of online stores. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This application is an exemplary network architecture that can be used to deploy and implement chatbot services;

[0027] Figure 2 A flowchart of an embodiment of the intelligent customer service response method of the present application;

[0028] Figure 3 This is a functional block diagram of the intelligent customer service response device of this application;

[0029] Figure 4 This is a schematic diagram of the structure of an intelligent customer service answering device used in this application. DETAILED DESCRIPTION

[0030] like Figure 1 As shown, the intelligent customer service response method of the present application covers a server 81 and a terminal device 80 in terms of network architecture. The server 81 undertakes the core task of executing the intelligent customer service response method to provide a chat robot for dialogue with the user, including processing user dialogue data, i.e., chat records generated by the dialogue between the user and the chat robot, obtaining user behavior data, and generating response information based on the user's latest question statement, and the response information includes existing products and recommended information of the online store. The server 81 realizes its core tasks by installing and running a computer program product implemented according to the intelligent customer service response method of the present application. The terminal device 80 is the interface for users to contact and use the chat robot. They are deployed through the corresponding application program, communicate with the server 81 over the network, and realize the sending and receiving of data.

[0031] In the application scenario of e-commerce platforms, the intelligent customer service response method of this application can customize the functions of chatbots according to different business needs to improve user experience and sales efficiency. For example, in online stores, chatbots can serve as virtual sales assistants to provide users with product consultation, recommendations, and purchase suggestions.

[0032] Taking the online store application scenario as an example, users interact with the chatbot through the application on the terminal device. The application has a friendly user interface and presents the chatbot as a virtual sales assistant for the online store, enhancing the user's shopping experience and desire for interaction. When the user asks a question or sends a message to the chatbot, the user question event is triggered, and these inputs are sent to the server as question statements through the terminal device. After the server receives the user's latest question statement, it uses the method of the present application to process the data to generate response information. The generated response information is then sent back to the terminal device and displayed to the user in the tone of a virtual sales assistant on the user interface, completing a complete interaction process.

[0033] In this application, the user's behavior data is used to describe and record the user's access behavior in the online store, which can be achieved by embedding points in the application or recording in the background program of the online store, so that the chatbot can provide more personalized and coherent services based on this. User behavior data can be defined by multiple behavior feature fields. User behavior analysis can fully and completely capture the user's shopping habits and preferences by including behavior data corresponding to behavior feature fields of multiple dimensions. Behavioral data can be stored in a behavior database for future use.

[0034] The behavioral feature fields used to define user behavior analysis can be set as needed to specify its data structure. The following are some exemplary behavioral feature fields: Purchase history: including the user's purchase history, product category and purchase frequency, etc. Browsing behavior: the user's browsing path, dwell time and page interaction in the online store, etc. Product preference: the user's preference for a specific product category or brand. User feedback: user's evaluation and feedback on products or services. Social influence: user interaction on social media and sharing behavior of products. Search history: user's search keywords and search frequency on e-commerce platforms. Geographic location: user's geographic location information helps to provide regionally relevant product recommendations. In addition, for each piece of behavioral data, the timestamp of the event is usually recorded.

[0035] The chat records generated by human-computer dialogue can be stored in the record database for future use. Chat records not only record every interaction between the user and the chatbot, including timestamps, conversation content, user emotions and other detailed information, but also through in-depth analysis of these data, the user's behavior patterns, preferences and needs can be revealed. These historical conversation data can help more accurately understand the user's intended products and their intention intensity in the process of generating response information, thereby generating more personalized and accurate response information. The detailed recording and analysis of chat records enable the chatbot to provide more considerate services in subsequent interactions and enhance the user experience. At the same time, it also provides e-commerce platforms with valuable user behavior data from another aspect, which is used to optimize product recommendations and improve sales efficiency.

[0036] See also Figure 2 According to an intelligent customer service response method provided by the present application, it can be implemented as a computer program product installed in a server and run, and after the computer program product runs, an intelligent customer service response system is constructed. In some embodiments thereof, the following steps are included:

[0037] Step S3100: Respond to a user question event and obtain the latest short chat record between the chat robot and the user, wherein the user question event includes the latest question statement of the user;

[0038] When a user interacts with a chatbot through a terminal device, every question or message sent by the user is considered a user question event. The system monitors and responds to these events in real time to obtain the latest conversation content with the user, which constitutes the so-called short chat record.

[0039] Short-term chat records refer to the conversation records between the user and the chatbot in the current session. They include the user's latest question and a limited number of previous interaction statements, such as the latest 10 interaction statements generated by the human-machine dialogue. These records can provide contextual information about the user's latest question, allowing the intelligent customer service system to more accurately understand the user's questions and intentions.

[0040] To achieve this step, the intelligent customer service system is responsible for receiving user question events submitted by the user's terminal device and parsing them to obtain the latest question statements, and then obtaining dialogue statements generated by multiple rounds of human-computer dialogues with limited data from the record database that stores historical chat records to form timely chat records.

[0041] Step S3200: determining semantically compatible historical behavior data from the user's behavior data according to the short chat records, and determining semantically compatible long chat records from the user's historical chat records;

[0042] The background information provided by short chat records is limited. In order to help deeply explore user intentions, data enhancement can be performed based on short chat records. By retrieving historical chat records with similar semantics to short chat records from the record database, and retrieving historical behavior data with similar semantics to short chat records from the behavior database, it helps to deeply explore user needs and preferences, so as to provide more accurate services.

[0043] Specifically, the system will extract historical behavior data that matches the semantics of the user's behavior data based on the latest short chat records. Historical behavior data mainly reflects the user's consumption pattern on the e-commerce platform, including but not limited to purchase history, browsing behavior, product preferences, etc. For example, if the user's historical behavior data shows that they often buy products of a certain brand, or have frequent browsing records for a certain category of products, this information will be used to assist in understanding the user's latest question statements, thereby inferring the user's possible purchase intention for related products.

[0044] At the same time, the system will also identify long chat records that match the user's semantics from the user's historical chat records. Long chat records contain long-term interaction information between the user and the chatbot, which can not only reveal the products that the user has been paying attention to for a long time, but also carry and reflect the user's expression patterns when purchasing products. For example, if the user often asks about discount information or usage methods of a specific product in past conversations, this information will help the system understand the context and intention of the user's questions.

[0045] In this application, the user expression patterns implied in the historical chat records play a positive role in the subsequent determination of the strength of intention. These records not only reflect the language habits and styles of users when interacting with chatbots, but also reveal their specific expressions during inquiries, discussions or purchases. For example, when asking about a product, a user may show a direct or indirect purchase intention, such as frequent use of certain keywords, phrases or question patterns, which can be used as a basis for judging the strength of the user's purchase intention. Through the expression patterns provided in the user's historical chat records, combined with the consumption patterns provided in the user's historical behavior data, the user's intended products and their intention strength can be more accurately identified, thereby providing product recommendations that are more in line with user needs and preferences when generating response information. This in-depth understanding of user expression patterns enables the intelligent customer service system to more accurately match the user's intentions, enhance the relevance and effectiveness of recommendations, and thus improve user satisfaction and purchase conversion rates.

[0046] Matching historical behavior data and long chat records based on short chat records can be achieved by vectorizing the short chat records, the behavior data in the behavior database, and the chat records in the record database to obtain their respective semantic vectors, and then calculating the data distance between the semantic vectors of the short chat records and the behavior data and chat records, and then filtering the data distance by a preset threshold. Usually, part of the behavior data whose data distance with the short chat records is higher than the preset threshold is obtained as the historical behavior data, and part of the chat records whose data distance with the short chat records is higher than the preset threshold is obtained as the long chat records.

[0047] Step S3300: merging the short chat records, long chat records and historical behavior data into question reference data, and determining the intended product that the user intends to consult and its corresponding intention strength according to the question reference data;

[0048] Short chat records provide contextual information of the user's current session, while long chat records reveal the user's preferences and behavior patterns in long-term interactions, implying the expression patterns corresponding to the user's intentions. Historical behavior data further supplements the user's consumption history and shopping habits on the e-commerce platform and implicitly provides consumption patterns. The combination of these three constitutes the reference data for questions, forming a multi-dimensional data set that provides rich information for understanding the user's complex needs. Based on this data set, the user's intended products and their intention strength can be more accurately determined.

[0049] Intended products refer to products that users show interest or demand in conversations, while the intensity of intent refers to the urgency or importance of users' interest or demand for these products. Regarding the intensity of intent, its level setting is defined based on the urgency or importance of users' interest and demand for specific products. For example, if a user mentions a certain product many times in a short chat record and shows an urgent willingness to buy, such as asking "Is this product in stock?" or "Can it be shipped today?", this high-intensity purchase intention can be identified as the highest intensity. On the contrary, if the user only asks for product information in general, such as "What are the pixels of this camera?", without showing the urgency of buying immediately, this intention can be identified as the second highest intensity or the lowest intensity. In addition, if the user's question is only a general understanding of the product, such as "What is the evaluation of this product?", this can be identified as the lowest intensity intention. This setting of the intensity of intent enables the system to provide more accurate and personalized product recommendations and services based on the different needs and purchasing intentions of users.

[0050] In one embodiment, a large language model can be used to implement intent recognition. The large language model is pre-trained to understand and process natural language data. By inputting question reference data into such a model, the large language model is instructed to identify the user's intended product and intent intensity based on the semantic information in the question reference data by providing relevant instructions. For example, if a user frequently mentions a certain product in the chat history, and historical behavior data shows that the user has browsed or purchased similar products, the large language model can combine this information to infer that the user has a high interest intensity in this product category.

[0051] In another embodiment, a self-trained deep learning model can also be used to replace the large language model to determine the user's intended products and their intention strength. This customized model can be trained according to specific business needs and user behavior characteristics to improve the accuracy of intent recognition. For example, the model can be trained to identify purchase signals for a specific category of products, or learn the language patterns of users in the purchase decision process.

[0052] Whether using a large language model or a customized deep learning model, the key is to be able to accurately extract the user's intended products and intent strength from the question reference data. This step is the basis for the intelligent customer service system to generate personalized response information, ensuring that the response information can meet the user's specific needs, thereby improving user experience and service quality. In this way, the intelligent customer service system can not only respond to the user's explicit needs, but also predict and meet the user's potential needs, achieving truly personalized services.

[0053] Step S3400, embed the question reference data, intended products and their intention strength into a preset answer prompt text, and generate answer information corresponding to the intention strength and matching the latest question statement according to the preset answer business logic of the answer prompt text, wherein the answer information includes the own products determined from the online store based on the intended products.

[0054] The integrated question reference data, including short chat records, long chat records and historical behavior data, as well as the intended products and intent strength identified therefrom, are embedded into the preset answer prompt text. By inputting the answer prompt text into the large language model and leveraging the large language model's ability to understand and process natural language data, personalized answer information can be generated based on these data.

[0055] Response prompt texts are predefined templates or frameworks designed to generate corresponding response information based on different intent products and intent strengths. These templates contain placeholders for inserting user-specific information, such as question reference data, intent products, and intent strengths, as well as instructions for guiding the large language model to understand and respond to the question reference data. After the user's intent product and intent strength are determined, this information, together with the question reference data, is embedded in the corresponding position of the response prompt text.

[0056] In this application, the instructions for guiding the large language model to understand and respond to the question reference data are implemented as a preset response business logic, which at least includes the following functions: one is to instruct the large language model to use the question parameter data as key background information; the second is to instruct the large language model to generate response information of a recommendation style that matches the intent strength according to a given intent strength; the third is to instruct the large language model to search for related products that are the same or similar to the intended product to form an associated product set, and then determine the own products in the associated product set from the product database of the online store. With the help of the response business logic, the large language model can generate response information corresponding to the latest question statement, which contains the own products corresponding to the intended product and provides a recommendation style that matches the user's intent strength.

[0057] To facilitate understanding, an example of response business logic is provided. Assume that the response business logic in the response prompt text contains the following instructions:

[0058] Background information prompt: "Consider the user's short chat records, long chat records and historical behavior data, identify the expression patterns and consumption patterns corresponding to the user's purchasing thinking, form an overall grasp of the user's consumption habits, and constrain the generation of response information accordingly."

[0059] Recommendation style prompt: "If the user shows high intent strength, the response should include direct product recommendations, calling the first recommendation style to encapsulate the response information; if the intent strength is medium, the response should include comparisons and suggestions, calling the second recommendation style to encapsulate the response information; if the intent strength is low, the response should provide information but not directly recommend, calling the third recommendation style to encapsulate the response information."

[0060] Product search prompt: "Search for related products that are the same or similar to the user's intended product, determine the related products from the online store database as the online store's own products, and the own products should be included in the response information."

[0061] The various recommended styles mentioned in the answer business logic can be preset in the form of a recommended style set and included in the answer prompt text for quick call by the large language model. An exemplary recommended style set is as follows:

[0062] The first recommendation style corresponds to the highest intent intensity. Example description: When a user asks "Is this camera available?", the response information includes: "Based on your needs, we recommend this camera you are interested in. It is currently in stock. Click here to view details." In the response information, the own product is constructed as prominent graphic information when displayed on the terminal device.

[0063] The second recommendation style corresponds to the second highest intent strength. Example description: When a user asks "How does this camera perform?", the response information includes: "The camera you are interested in has excellent performance and is suitable for professional photography. You can also consider this more cost-effective camera." In the response information, the own product is constructed as non-prominent graphic information when displayed on the terminal device, but derivatives related to the own product are additionally displayed, such as displaying one or more photographic works corresponding to the camera.

[0064] The third recommended style corresponds to the lowest intent strength. Example description: When the user only asks "What brands of cameras are there?", the response information includes: "We provide cameras from multiple brands, including Brand A, Brand B, etc. You can learn more about the characteristics of each brand according to your needs." In the response information, a simple list of our own products is included, and a manual seat interface is added.

[0065] The design of the answer business logic and the recommended style set enables the large language model to generate appropriate answer information based on different conversation situations and user intent strength. When the large language model receives the answer prompt text, it first parses the instructions in the answer business logic, which instruct the model how to understand and process the question reference data and generate answer information accordingly.

[0066] Taking the first recommendation style as an example, when the user's intent strength is identified as the highest level, the model will call the first recommendation style to generate a response. This style directly and prominently recommends own products, because the high intent strength indicates that the user is likely to make a purchase. For example, if the user asks "Is this camera in stock?" The model will understand that this is a highly relevant purchase signal and directly recommend the camera in the response, while providing a purchase link or further details, allowing the user to take action quickly.

[0067] For the second recommendation style, which is suitable for situations with the second highest intent strength, the model will provide comparisons and suggestions in the response. This approach not only recommends the products that the user asked about, but also recommends other related or similar products to increase the user's choice range and the possibility of purchase. For example, if the user is curious about the performance of a certain camera, the model will not only provide performance information of the camera, but also recommend other cost-effective cameras, and may show some photography works to enhance the attractiveness of the recommendation.

[0068] The third recommendation style is suitable for the situation with the lowest intention strength, when the user's purchase intention is not yet clear, and the model's goal is to provide information rather than directly recommend products. For example, when a user asks about camera brands, the model will list multiple brands and provide ways to learn more, while providing a human agent interface so that users can get more personalized help when needed.

[0069] In this process, the large language model generates response information that matches the user's intent strength based on the instructions in the response business logic and the preset styles in the recommended style set. This generation process is dynamic, and the model will flexibly adjust the response content based on the user's specific situation and conversation context to ensure the relevance and effectiveness of the response information.

[0070] In this way, the large language model can not only understand the user's intentions, but also provide personalized product recommendations based on these intentions, thereby improving user experience and satisfaction. This intelligent response generation method enables the intelligent customer service system to more accurately meet user needs, improve service efficiency and sales conversion rate.

[0071] It is not difficult to understand from the above embodiments that the present application has many technical advantages over the traditional technology, including but not limited to:

[0072] First, this application can more accurately capture the user's true intentions and needs by comprehensively analyzing the user's short chat records, historical behavior data, and long chat records. This multi-dimensional data fusion method effectively overcomes the problem that the user portrait labels in the prior art are too abstract and cannot provide comprehensive semantic information, so that the generated response information can more accurately correspond to the user's true intentions, improving service quality and user satisfaction.

[0073] Secondly, this application is more sophisticated and flexible in terms of product recommendation logic. It not only takes into account the user's behavioral data and the semantic similarity between the chat records and the question statements, but also comprehensively considers the short-term chat records and long-term chat records in the human-computer dialogue process. It also further clarifies the user's intended products and their intention strength, and makes corresponding recommendations in the response information accordingly, providing different recommendation schemes for different intention strengths. This recommendation method based on intention strength can better stimulate users' willingness to buy and effectively improve the sales conversion rate of online stores.

[0074] In addition, the application guides the generation of response information through the response business logic in the preset response prompt text, making the generation process of response information more intelligent. The application of this intelligent decision-making logic significantly improves the flexibility of recommending products and improves the recommendation results, thereby more effectively promoting the sales goals of online stores.

[0075] Based on any embodiment of the method of the present application, determining the intended product that the user intends to consult and its corresponding intention strength according to the question reference data includes:

[0076] Step S3310: embedding the question reference data into the intention recognition prompt text, wherein the intention recognition prompt text includes an intention recognition control instruction for instructing to refer to the short chat records, long chat records and historical behavior data to recognize the user's latest intended product and its intention strength;

[0077] The collected reference data for questions, including short chat records, long chat records, and historical behavior data, are embedded in a specially designed intent recognition prompt text, so that this prompt text contains not only the user's conversation data, but also intent recognition control instructions. These control instructions are key information used to guide the large language model on how to analyze and process this data. The control instructions tell the model the type of data that needs special attention, such as the user's purchase history, browsing behavior, or specific keywords and phrases used by the user in the conversation. These instructions ensure that the model can extract the most relevant information from a large amount of data, thereby accurately identifying the user's intended products and intent strength.

[0078] An exemplary intent recognition prompt text may be as follows: "Analyze the user's latest question statement in the following short chat record, and combine the historical behavior data provided by the user and the long chat record to determine the specific product the user is asking about and the strength of his / her purchase intention. Control instructions: Comprehensively examine the user's expression pattern of his / her purchasing thinking and consumption pattern of purchased products through long chat records and user behavior data, giving priority to the user's search and purchase history in the past week, paying special attention to the brand and model of products mentioned by the user, and any keywords related to purchase decisions, such as 'price', 'discount', 'purchase', etc." Such prompt text provides a clear analysis path for the large language model, enabling it to accurately extract the user's intended product and intent strength from the question reference data.

[0079] Step S3320: input the intention recognition prompt text into the large language model to control the large language model to determine the user's latest intended product and its intention strength according to the intention recognition control instruction and with reference to the time information and text semantics provided by the question reference data.

[0080] The large language model identifies the control instructions in the prompt text according to the intent, analyzes the time information and text semantics provided by the timestamp in the question reference data, and thus determines the user's latest intended product and intent strength.

[0081] The large language model identifies keywords and phrases in user conversations, analyzes their correlation with the user's historical behavior data, and the changing trends of these behavior data over time. For example, if a user frequently mentions a product of a specific brand in a chat history, and historical behavior data shows that the user has browsed or purchased related products of that brand, the large language model will combine this information and infer that the user may have a high interest in or purchase intention for the brand's products.

[0082] In this way, the large language model can generate an accurate user intent profile, including the user's intended products and intent strength. This profile provides a solid foundation for the subsequent response generation step, ensuring the relevance and personalization of the response information, thereby improving user satisfaction and service quality. This intent recognition method based on the large language model not only improves the accuracy of recognition, but also makes the entire response generation process more efficient and intelligent.

[0083] On the basis of any embodiment of the method of the present application, the question reference data, the intended product and its intention strength are embedded in a preset answer prompt text, and according to the answer business logic preset in the answer prompt text, answer information corresponding to the intention strength and matching the latest question statement is generated, and the answer business logic instructs the large language model to execute the following business process:

[0084] Step S3410: searching online for a set of related products that match the comprehensive semantic information based on the intended product and the comprehensive semantic information of the question reference data;

[0085] In this step, the user's intended product and question reference data are first analyzed, and semantic information that helps to strengthen the description of the intended product is extracted from the question reference data. This information can include the specific product features, brand, model, purpose, etc. that the user inquired about, as well as any other details related to the intended product. For example, if the user's intended product is a specific smartphone, then the key semantic information may include the brand of the phone, the technical parameters that the user cares about (such as camera pixels, processor speed, etc.), and the user's usage scenarios (such as photography, games, etc.).

[0086] After obtaining key semantic information, online searches can be performed on the Internet or all e-commerce platforms. The purpose of the search is to find related products that are the same or similar to the user's intended product. The result of the online search is a set of related products, which contains all products that semantically match the user's intended product. These products may come from different merchants, brands, or categories, but they are all related to the user's original query. When obtaining these related products, the description information of the related products is usually also included in the related product set, including but not limited to product titles, product images, product attributes, product details descriptions, and other information. The related product set provides a wealth of candidate products for subsequent steps, enabling the system to provide users with a variety of product choices.

[0087] For example, if a user asks about a specific camera model, the system can search and return other camera models of the same brand, similar products from competitors, or photography-related accessories such as lenses, tripods, etc. Such a collection of related products not only helps users find the product they originally asked about, but may also lead them to discover new products that may better meet their needs.

[0088] Step S3420: matching the own products stored in the product database of the online store according to the description information of the related products in the related product set;

[0089] By comparing the description information provided by the associated product set with the product database of the online store, the online store's own products can be determined. The product database of the online store is a system that contains detailed information on all the products available for sale in the online store, which may include product name, model, specification, inventory status, price, etc. Further, each product in the associated product set is matched with an entry in the database through a text matching algorithm, database query or machine learning model.

[0090] The matching process can cover multiple levels of operations. For example, first, check whether the items in the associated product set have corresponding entries in the product database. If so, the inventory status of these items can be further verified to ensure that the recommended items are available for purchase. In addition, other factors such as the price competitiveness of the items, user reviews, historical sales data, etc. can also be considered to evaluate the suitability of these items as recommendation candidates.

[0091] Step S3430: Adapt the intention strength to determine the corresponding preset recommendation style in the response business logic, and construct response information corresponding to the latest question semantics according to the recommendation style.

[0092] In this step, according to the response business logic and the strength of the user's intention, a preset recommendation style is selected to construct the response information. As mentioned above, the recommendation style can be divided into several levels according to the strength of the intention, generally two or more, and each level corresponds to a different recommendation strategy and expression. For example, for high intention strength, the recommendation style can be a direct and clear product recommendation; for medium intention strength, the recommendation style can be to provide comparisons and suggestions; and for low intention strength, the recommendation style can be to provide information but not directly recommend products.

[0093] Furthermore, the selected recommendation style is applied to the construction of the response information. This includes embedding the matched own product information into the corresponding position of the response prompt text, and adjusting the tone and content of the response information according to the recommendation style. For example, if the user's intention strength is high, a response containing a direct purchase link and promotional information can be constructed; if the intention strength is low, a response containing a product information list and a further learning option can be provided.

[0094] In addition, you can also consider how to present your own products to users in the most attractive way. This may include using attractive product images, highlighting the key features of the product, or providing user reviews and ratings. The goal is to create a response that is both in line with user intent and attractive to increase user purchase intention and satisfaction.

[0095] The technical advantage of this embodiment is that it can provide a highly personalized and accurate answering service, which is mainly due to its comprehensive use of users' short chat records, long chat records and historical behavior data to determine the intended products and the strength of the intention, and then search and match the associated product set and the own products online, ensuring the use of comprehensive associated products to target the own products in the online store. In this way, not only can the user's immediate needs be understood, but also the user's potential needs can be predicted and met, providing a richer and more diverse selection of goods. In addition, different recommendation styles are adapted according to the strength of the intention, so that the response information is more in line with the user's purchasing intention, thereby improving the user experience and increasing the conversion rate. This intelligent response generation method enables the intelligent customer service system to more accurately meet the needs of users and improve service efficiency and sales results.

[0096] Based on any embodiment of the method of the present application, adapting the intention strength to determine a corresponding preset recommendation style in the response business logic includes any one or more of the following steps:

[0097] Step S3431: when the intention strength is the highest strength, determine a first recommendation style, the first recommendation style indicating that the self-owned product is directly and prominently recommended in the corresponding response information;

[0098] As mentioned above, when the user's intention strength is identified as the highest strength, the first recommendation style will be used to indicate how to construct the response information. The purpose of this style is to directly and prominently recommend the own products in the response information, because the high intention strength indicates that the user is likely to make a purchase. In this case, a direct recommendation style can be determined, which can highlight the key information of the product, such as discounts, inventory status, purchase links, etc., to encourage users to make a quick purchase decision. The response information of this style is designed to be more direct and urgent to match the user's high purchase intention.

[0099] Step S3432: when the intention strength is the second highest strength, determine a second recommendation style, the second recommendation style indicating adding associated recommendation information for enhancing the user's intention to purchase the owned product to the corresponding response information;

[0100] For the second highest intensity intention, the second recommendation style is determined accordingly. This style aims to enhance the user's intention to purchase the own product by adding related recommendation information to provide more choices and information. For example, other products similar or related to the product inquired by the user can be recommended, or some additional preferential information such as bundling or discounts can be provided. This style of response information aims to increase the user's purchase probability by providing additional choices and incentives.

[0101] Step S3433: When the intention strength is the lowest strength, determine a third recommendation style, the third recommendation style indicating providing the owned product in the corresponding response information and adding a manual agent dialogue interface.

[0102] When the user's intent strength is identified as the lowest intensity, the third recommended style is adopted. This style is more conservative. It provides its own product information while adding a human agent dialogue interface. This is because the lowest intensity intent indicates that the user may still be in the information collection stage and is not ready to buy. Therefore, product information can be provided, and an option can also be provided to allow users to communicate further with human customer service to obtain more information or resolve questions. This style of response information is designed to keep users engaged while providing them with a way to interact further.

[0103] Through this meticulous adaptation, the intelligent customer service system can provide customized responses based on different intent strengths, thereby improving user satisfaction and purchase conversion rates. This flexible recommendation mechanism is a notable feature of the present application method, which enables the intelligent customer service system to more accurately meet user needs and provide appropriate support at different purchase stages.

[0104] Based on any embodiment of the method of the present application, after generating the response information corresponding to the intention strength and matching the latest question statement, it includes:

[0105] Step S3500: obtaining the corresponding recommendation behavior data of the users accessing the own products in the online store after the products are constructed into the response information in a determined recommendation style for recommendation;

[0106] After the user receives the response information customized according to the strength of his / her intention, the user's access behavior to the recommended products is monitored and recorded. Such behavior data includes, but is not limited to, the frequency of users clicking on the recommended links of their own products, the time they stay on the product page, the number of products added to the shopping cart, and the final purchase conversion rate. In some embodiments, such data can be collected by implanting tracking codes, using website analysis tools, or integrating the built-in analysis function of the e-commerce platform.

[0107] For example, if the response information recommends a smartphone and provides a direct purchase link, the corresponding record is how many users clicked on the link and whether these clicks led to actual purchases. If the recommendation style is to provide comparisons and suggestions, the corresponding analysis is whether users browsed the recommended products based on these suggestions, and whether these recommendations promoted the user's purchase decision.

[0108] The purpose of collecting this data is to evaluate the effectiveness of different recommendation styles and provide a basis for subsequent optimization. By analyzing the user's click-through rate, dwell time, and conversion rate, we can understand which recommendation styles are most attractive to users and adjust the recommendation strategy accordingly.

[0109] In addition, machine learning algorithms can be used to conduct in-depth analysis of these behavioral data to identify patterns and trends in user behavior. These analysis results can help the system predict users' purchasing intentions and adjust recommendation strategies in real time to improve the relevance and attractiveness of recommendations.

[0110] Step S3600: Calculate the recommendation effectiveness data of each recommendation style according to the recommendation behavior data;

[0111] The statistics of recommendation effectiveness data may include the calculation of one or more indicators, including but not limited to click-through rate (CTR), conversion rate (CVR), page dwell time, add-to-cart rate, etc. The click-through rate is calculated by dividing the number of users who click on the recommended link of the own product by the total number of users who see the recommended link. This indicator reflects the degree to which the recommendation attracts the user's attention. The conversion rate is calculated by dividing the number of users who complete the purchase through the recommended link of the own product by the total number of users who click on the recommended link. This indicator directly reflects the ability of the recommendation to promote purchases. The page dwell time reflects the user's interest in the own product, while the add-to-cart rate is an intermediate indicator of the user's interest in the recommended own product and taking action.

[0112] In order to obtain the recommendation effectiveness data for each recommendation style, these indicator data of all users are classified according to the recommendation style, and the average indicator value of each style is calculated. For example, if the first recommendation style is a direct product recommendation, the average click rate and conversion rate of all recommendation links using this style are calculated. Similarly, for the second and third recommendation styles, their recommendation effectiveness data can also be calculated separately.

[0113] When calculating these averages, a weighted algorithm can be used to consider the importance of different users or different products. For example, if a user's historical purchase data shows that they are more likely to make high-value purchases, then the user's click and conversion data can be given a higher weight. Similarly, if a certain product is a key promotional product of the store, then the recommendation effectiveness data of this product can also be given a higher weight.

[0114] In this way, various factors can be comprehensively considered to obtain the recommendation effectiveness data of each recommended style. This recommendation effectiveness data is the result of weighted fusion of multiple preset indicators. These data not only reflect the overall performance of each style, but also help the system identify which styles are more effective in specific situations. For example, the system may find that when the user's intention intensity is high, direct product recommendations (the first recommended style) are most effective, while when the user's intention intensity is low, providing information but not directly recommending products (the third recommended style) can attract users to explore further.

[0115] Step S3700: determine whether the recommendation effectiveness data meets a preset condition. If the preset condition is not met, readjust the corresponding recommendation style in the response prompt text.

[0116] The recommendation effectiveness data of each recommendation style not only reflects the user's response to the current recommendation style, but also provides a basis for whether the recommendation strategy needs to be adjusted.

[0117] When the recommendation performance data does not meet the preset conditions, for example, when it is below the preset threshold, the recommendation style can be readjusted. This process can be achieved in many ways, one of which is to leverage the capabilities of large language models. Large language models can understand and generate natural language text. Based on the feedback from the recommendation performance data, the model can adjust the language style, content, and structure of the answer prompt text in order to improve the user's response.

[0118] For example, if the conversion rate of the first recommended style is lower than expected, the large language model can add more elements to the response prompt text to attract users' attention, such as discount information, user reviews, or the uniqueness of the product. Correspondingly, multiple versions of the response prompt text are generated, and the effects of different versions are compared through A / B testing and other methods, and the best version is selected for actual response.

[0119] In addition to using a large language model, other embodiments can also be used for adjustment. For example, the parameters in the recommendation logic can be automatically adjusted according to the trend of the recommendation effectiveness data, such as changing the weight distribution in the recommendation algorithm, or adjusting the triggering conditions of the recommendation style. In addition, user feedback information can be introduced as a supplementary indicator of recommendation effectiveness, and the recommendation style can be further optimized by analyzing the direct feedback of users.

[0120] In practice, multiple methods can be combined to readjust the recommendation style. For example, a large language model can be used to generate multiple variants of the response prompt text, and then a machine learning algorithm can be used to analyze the correlation between these variants and user behavior data to determine the most effective response prompt text. This method combines natural language processing and data-driven decision-making, allowing the system to respond to user intentions more accurately and provide recommendations that better meet user needs.

[0121] The significant technical advantage of this embodiment is that it can achieve real-time monitoring and dynamic adjustment of recommendation results by dynamically adjusting the recommendation style, thereby significantly improving the adaptability and effectiveness of the recommendation mechanism of the intelligent customer service system. By collecting and analyzing the user's behavioral data on recommended products, the actual effectiveness of different recommendation styles can be accurately measured and optimized accordingly. This data-driven approach enables the recommendation mechanism to not only respond to users' immediate needs, but also predict and adapt to changes in user behavior and provide more personalized recommendations.

[0122] Compared with other embodiments, this embodiment provides a direct feedback mechanism for optimizing the recommendation mechanism by introducing statistics and evaluation of recommendation effectiveness data. This means that it is possible to automatically identify which recommendation styles are most effective and which need to be improved, thereby achieving more accurate marketing and higher user satisfaction. In addition, this embodiment also utilizes the capabilities of a large language model to dynamically adjust the response prompt text through natural language processing technology, which not only improves the attractiveness of the recommendation, but also enhances the flexibility and response speed of the system.

[0123] It can be seen that the technical advantage of this embodiment is that it comprehensively uses data analysis, machine learning and natural language processing technologies to form a closed-loop recommendation optimization process. This process not only improves the accuracy and relevance of recommendations, but also provides users with a richer and more attractive shopping experience, which is of great significance for improving the competitiveness and market share of e-commerce platforms.

[0124] Based on any embodiment of the method of the present application, the corresponding recommended style in the answer prompt text is readjusted, including:

[0125] Step S3710: embed each recommended style and its corresponding recommended effectiveness data into a style prompt text, wherein the style prompt text includes a style adjustment instruction, wherein the style adjustment instruction indicates a target recommended style that needs to be adjusted to improve its recommended effectiveness data;

[0126] The style hint text contains the specific content of the recommended styles, as well as embedded recommendation performance data and style adjustment instructions about these recommended styles. The recommended performance data and style adjustment instructions provide direct feedback on the performance of each recommended style, indicating which recommended styles need to be adjusted to improve their recommendation effectiveness. For example, if any one or more of a recommended style's click-through rate, conversion rate, page dwell time, add-to-cart rate, and specific recommendation performance data are lower than expected, by embedding these data feedbacks into the style hint text and attaching a style adjustment instruction, it can be clearly pointed out that this style needs to be optimized to improve its effectiveness.

[0127] Style adjustment instructions are used in the style prompt text to explain what specific adjustments need to be made to the recommended style. Style adjustment instructions can be instructions to add or subtract certain content, change the recommended language style, adjust the location or format of the recommendation, or introduce new elements such as user reviews, special offers, etc. The specific content of the style adjustment instructions will be determined based on the analysis results of the recommendation effectiveness data.

[0128] For example, in a specific embodiment, if the recommendation style containing user evaluation has a higher conversion rate, the style adjustment instruction may indicate to add user evaluation content to all recommendation styles. Such instructions will directly guide the large language model to generate a new recommendation style version for subsequent testing and optimization.

[0129] By embedding the recommended style and its effectiveness data into the style prompt text and attaching clear style adjustment instructions, we can ensure that each optimization is based on actual recommendation effectiveness data feedback and has a clear direction. This method makes the optimization process of the recommendation mechanism both data-driven and goal-oriented, thereby improving the efficiency and effectiveness of optimization.

[0130] Step S3720: input the style prompt text into a large language model to generate multiple versions of the target recommended style according to the style adjustment instruction;

[0131] In this step, the large language model receives a style prompt text containing recommended styles, recommended effectiveness data, and style adjustment instructions. The model uses its advanced text understanding and generation capabilities to adjust the existing recommended styles according to the specific requirements of the style adjustment instructions and generates multiple new recommended style versions. These versions reflect different adjustment directions, such as changing the recommended language style, adding or reducing content, adjusting the format, etc., in response to the style adjustment instructions in the style prompt text.

[0132] For example, if the style adjustment instructions indicate that user reviews need to be added to improve conversion rates, the large language model will create a recommended style version that includes user reviews. After applying these recommended style versions, the model will not only extract user review information related to the recommended own products from the database, but also integrate these reviews into the recommended text in a natural and attractive way, generating multiple different text versions, each slightly different, to test which method is most effective.

[0133] In addition, the large language model can also make more complex adjustments based on style adjustment instructions, such as generating new recommendation styles based on the latest market trends, user behavior patterns, or special promotions. The model can understand and apply various language styles and expressions to ensure that the generated recommendation styles are both consistent with business logic and attractive and persuasive.

[0134] In the style prompt text, you can constrain the number of versions of generated recommended styles and the expression method of generated recommended styles. For example, you can define the recommended style in Json format, so that when you embed the recommended style into the response prompt text later, you can automatically generate the corresponding recommended style content and include it in the response information based on the definition of the recommended style. For example, you can load user evaluation content of your own products from the product database of an online store into the response information based on the definition of the recommended style.

[0135] Step S3730: put the generated multiple versions into the response business logic of the response prompt text for comparison testing, and replace the target version with the highest recommended effectiveness data in the comparison test results with the determined version in the response business logic.

[0136] The purpose of comparative testing is to evaluate the effects of different recommendation style versions and select the best version to improve user experience and recommendation effectiveness. Through comparative testing, we can understand the performance of each recommendation style version in actual application, including the recommendation effectiveness data composed of key indicators such as user interaction, click-through rate, conversion rate, etc., so as to provide data support for selecting the best recommendation style.

[0137] When implementing comparative testing, multiple recommendation style versions generated by the large language model are actually applied to user interactions. This process can be achieved through A / B testing or multi-version testing, that is, during the conversation between the user and the chatbot, different recommendation style versions are randomly or according to specific rules to the user group, and then user behavior data is collected, and the recommendation effectiveness data of each recommendation style version is determined based on the user behavior data. For example, if three different recommendation style versions are generated, each version differs in language style, content length, or user evaluation display method, the system will monitor the specific reactions of users to these versions, including the proportion of clicks on the recommendation link, the duration of stay on the product page, the number of products added to the shopping cart, etc., and accordingly, the recommendation effectiveness data is determined by referring to the method disclosed in the previous article.

[0138] In the specific implementation of comparative testing, diversion technology can be used to evenly distribute user traffic to different recommended style versions to ensure that each version has a fair testing opportunity. By recording the user behavior data of each version and using statistical analysis methods to evaluate the performance of each version, the recommendation effectiveness data of each version can be determined. For example, if a certain version of the recommended style shows a higher user click-through rate and conversion rate, these recommendation effectiveness data can be used alone or weighted as the basis for selecting this version as the final recommended style.

[0139] In addition, in one embodiment, a large amount of collected user behavior data can be deeply analyzed in combination with a machine learning algorithm to identify the preferences of different user groups for different recommendation styles. This analysis can help the system adjust the recommendation strategy more accurately and apply the response information of different groups of users to meet the needs and preferences of different users.

[0140] Through this comparative testing and data analysis method, we can continuously iterate and optimize the recommendation style to ensure that the recommended content is always consistent with user needs, thereby improving the accuracy of the recommendation mechanism and user satisfaction. Finally, we select the recommendation style version with the highest recommendation effectiveness data, determine it as the final version in the response business logic, and apply it to actual user interactions.

[0141] The outstanding technical advantage of this embodiment is that it can achieve refined management and dynamic optimization of recommendation styles. By embedding the recommended style and its effectiveness data into the style prompt text, and using style adjustment instructions to guide the large language model to generate multiple recommended style versions, this embodiment can ensure that each recommendation is based on the most cutting-edge data analysis and user feedback. This method not only improves the personalization of recommendations, but also enhances the adaptability and flexibility of the system, enabling it to quickly respond to market changes and the evolution of user needs.

[0142] Compared with other embodiments, this embodiment can accurately evaluate the actual effects of different recommendation style versions and select the optimal version through comparative testing and data analysis. This empirical method ensures that the optimization decision of the recommendation mechanism has higher reliability and effectiveness. In addition, this embodiment can also identify the preferences of different user groups based on in-depth analysis of user behavior data and machine learning algorithms, and achieve more accurate personalized recommendations.

[0143] In general, the technical advantage of this embodiment is that it uses data-driven decision-making, advanced natural language processing technology and machine learning algorithms to form a closed-loop recommendation optimization process. This process not only improves the accuracy and relevance of recommendations, but also provides users with a richer and more attractive shopping experience, which is of great significance for improving the competitiveness and market share of e-commerce platforms. Through this continuous iteration and optimization, this embodiment can ensure that the recommendation mechanism is always in the best state and provide users with the most valuable suggestions and information.

[0144] Based on any embodiment of the method of the present application, the method further includes:

[0145] Step S1100: asynchronously updating the behavior data generated by the user in the online store, formatting it and storing it in a behavior database, so that the behavior data in the behavior database includes a timestamp for providing time information and behavior event data for describing corresponding behavior events;

[0146] During the operation of the intelligent customer service system, continuously updating and maintaining user behavior data can not only help understand user needs and preferences, but also lay the foundation for providing personalized services. Specifically, the behavior data generated by users in online stores can be updated asynchronously, and the data can be formatted and stored in the behavior database. The update of behavior data is a background process, which will not interfere with the real-time interaction between users and chatbots.

[0147] Behavioral data includes various activity records of users on e-commerce platforms, such as browsing history, search queries, purchase behavior, product reviews and collections, etc. Each piece of behavioral data is accompanied by a timestamp and behavioral event data. The timestamp records the specific time when the behavior occurred, which helps to analyze user behavior patterns and trends. At the same time, behavioral event data describes the user's specific behavior in detail, such as "User A purchased a smartphone on November 27, 2023" or "User B browsed the details page of a laptop on November 26, 2023."

[0148] In one embodiment, these behavioral event data can be formatted to ensure that they can be effectively stored and queried. The formatting process can include converting the data into a unified data structure, encoding the data to reduce storage space or improve query efficiency, and cleaning the data to remove invalid or erroneous records. The stored behavioral data will be used for subsequent analysis and processing. For example, when a user interacts with a chatbot, relevant behavioral data can be quickly retrieved to provide more personalized responses and recommendations.

[0149] For example, if a user asks about a camera in a chat, the behavior database can be quickly queried to find out whether the user has ever browsed or purchased a camera, or whether the user has shown interest in photography-related products. This information will help generate more relevant and accurate response information and improve the user experience. In this way, real-time monitoring and analysis of user behavior data can be achieved to provide support for personalized services.

[0150] Step S1200, asynchronously updating the chat records generated by the conversation between the user and the chat robot, and formatting and storing the chat records in the record database, so that the chat records in the record database include a timestamp for providing time information and multiple rounds of conversations generated by the interaction between the user and the chat robot;

[0151] The continuous updating and storage of chat records helps to understand user needs and optimize the conversation experience. The conversation records between the user and the chatbot can be updated asynchronously and stored in the record database after formatting. This process is also carried out in the background and will not affect the real-time communication between the user and the chatbot.

[0152] Chat records include all information generated during the interaction between the user and the chatbot, such as questions, responses, and any form of feedback. Each chat record is accompanied by a timestamp to record the specific time when the conversation occurred, so that the conversation process can be tracked and user behavior patterns can be analyzed. At the same time, the chat record data records the content of the conversation in detail, including the user's question statements and the robot's response information, as well as the context information of the conversation.

[0153] In one embodiment, the formatting of the chat record includes converting the conversation content into a unified data structure for easy storage and query, including converting the text into a specific data format, assigning a unique conversation identifier, and ensuring that all messages are arranged in chronological order. In addition, formatting can also include data cleaning operations, such as removing invalid or redundant information, and ensuring the consistency and accuracy of the data.

[0154] The stored chat records will be used for subsequent analysis and processing. For example, when the user interacts with the chatbot again, the previous conversation records can be quickly retrieved to provide contextual answers and suggestions. For example, if the user mentioned their preference for a certain product in a previous conversation, this information can be cited in subsequent conversations to provide users with more personalized services.

[0155] In addition, the storage of chat records also helps the intelligent customer service system to learn and optimize itself. By analyzing the interaction between users and chatbots, common problems and user demand patterns can be identified, thereby improving response strategies and enhancing user experience. For example, if analysis shows that users often ask for specific information about a product, the chatbot's knowledge base can be optimized to provide relevant information more quickly.

[0156] Step S1300: vectorize the behavior data and chat records for semantic matching based on the short chat records between the user and the chat robot to determine the associated historical behavior data and the historical chat records.

[0157] Vectorizing behavior data and chat records is a key step in achieving accurate semantic matching. Vectorization is the process of converting text data into numerical vectors that can represent the semantic features of the text and are used to calculate the similarity between texts. In this way, historical behavior data and historical chat records that match the semantics can be found based on the short chat records between the user and the chatbot.

[0158] In one embodiment, the behavior data and chat records are first preprocessed, including natural language processing steps such as word segmentation, stop word removal, and stem extraction to extract key information. Then, vectorization techniques such as bag-of-words model, TF-IDF or Word2Vec are used to convert the text into fixed-length numerical vectors. These vectors capture the semantic information of the text and can be used for subsequent calculations and comparisons.

[0159] In another embodiment, deep semantic vectors of behavior data and chat records may be directly extracted with the help of a deep semantic model to achieve vectorization.

[0160] It can be seen that in different embodiments, different vectorization techniques and similarity calculation methods can be used. For example, some embodiments can use a deep learning-based vectorization model, such as BERT or GPT, to capture more complex semantic relationships. Other embodiments can combine multiple vectorization techniques to improve the accuracy and robustness of matching.

[0161] After determining the vectors of behavior data and chat records, the vectors of short-term chat records can also be determined by the same logic. Based on this, data distance algorithms such as cosine similarity and Qu's distance algorithm can be used to calculate the similarity between the vectors of short-term chat records and the vectors of behavior data and chat records, respectively, and the most relevant data can be determined based on the similarity scores, and some behavior data with higher similarity scores can be selected to form historical behavior data, and some historical chat records can be selected to form long-term chat records. High scores indicate that they are most semantically relevant to the current user's short-term chat records, so they can be used to provide more personalized and accurate responses.

[0162] The outstanding technical advantage of this embodiment is that it can realize the continuous update and in-depth analysis of user behavior data and chat records, thereby providing strong data support for the intelligent customer service system. By asynchronously updating the user's behavior data in the online store and the conversation record with the chatbot, the data can be kept up to date without affecting the real-time interaction between the user and the system. This continuous data update mechanism ensures that the system can generate responses based on the latest user behavior and interaction history, improving the timeliness and relevance of the responses.

[0163] In addition, this embodiment achieves accurate semantic matching by vectorizing behavior data and chat records. Vectorization technology enables the system to convert text data into numerical vectors and then calculate the similarity between texts. In this way, it can not only identify the user behavior and historical chat records most relevant to the current conversation, but also provide more personalized services.

[0164] Compared with other embodiments, the advantage of this embodiment is that it can combine multiple vectorization technologies and similarity calculation methods to improve the accuracy and robustness of matching. This flexibility enables the system to adapt to different application scenarios and needs, whether it is a vectorization model based on traditional machine learning or an advanced model based on deep learning, it can flexibly select the most appropriate method.

[0165] Ultimately, the technical advantage of this embodiment is that it can provide a more accurate and personalized user experience. Through continuous data updates and in-depth semantic analysis, it can better understand user needs, predict user behavior, and provide timely, relevant and useful responses. This data-driven approach not only improves the efficiency of the intelligent customer service system, but also enhances user satisfaction and trust in the system.

[0166] See also Figure 3According to one aspect of the present application, an intelligent customer service answering device is provided, including a question response module 3100, a data matching module 3200, an intention recognition module 3300, and an answer generation module 3400, wherein the question response module 3100 is configured to respond to a user question event and obtain the latest short-term chat record between the chat robot and the user, wherein the user question event includes the latest question statement of the user; the data matching module 3200 is configured to determine semantically compatible historical behavior data from the user's behavior data based on the short-term chat record, and determine semantically compatible historical behavior data from the user's historical chat record. long chat records; the intention recognition module 3300 is configured to merge the short chat records, long chat records and historical behavior data into question reference data, and determine the intended product that the user intends to consult and its corresponding intention strength according to the question reference data; the answer generation module 3400 is configured to embed the question reference data, the intended product and its intention strength into a preset answer prompt text, and generate answer information corresponding to the intention strength and matching the latest question statement according to the answer business logic preset in the answer prompt text, wherein the answer information includes the own product determined from the online store according to the intended product.

[0167] Based on any embodiment of the device of the present application, the intention recognition module 3300 includes: an identification prompt construction module, which is configured to embed the question reference data into the intention recognition prompt text, and the intention recognition prompt text contains an intention recognition control instruction, which is used to instruct to refer to the short chat records, long chat records and historical behavior data to identify the user's latest intention products and their intention strength; an intention recognition execution module, which is configured to input the intention recognition prompt text into a large language model to control the large language model to determine the user's latest intention products and their intention strength according to the intention recognition control instruction and with reference to the time information and text semantics provided by the question reference data.

[0168] Based on any embodiment of the device of the present application, in the response generation module 3400, the response business logic instructs the large language model to execute the following business process to construct the following modules: an online search module, configured to search online for a set of related products that match the comprehensive semantic information of the intended product and the question reference data; a product determination module, configured to match the owned products stored in the product database of the online store based on the description information of the related products in the related product set; a style adaptation module, configured to adapt the intention strength to determine the corresponding preset recommended style in the response business logic, and construct response information corresponding to the latest question semantics according to the recommended style.

[0169] Based on any embodiment of the device of the present application, the style adaptation module includes any one or more of the following modules: a first recommendation module, configured to determine a first recommendation style when the intention strength is the highest intensity, and the first recommendation style indicates that the owned product is directly and prominently recommended in the corresponding response information; a second recommendation module, configured to determine a second recommendation style when the intention strength is the second highest intensity, and the second recommendation style indicates that related recommendation information for enhancing the user's intention to purchase the owned product is added to the corresponding response information; a third recommendation module, configured to determine a third recommendation style when the intention strength is the lowest intensity, and the third recommendation style indicates that the owned product is provided in the corresponding response information and a manual agent dialogue interface is added.

[0170] Based on any embodiment of the device of the present application, the device also includes: a recommendation behavior monitoring module, configured to obtain the corresponding recommendation behavior data of the own products in the online store after being constructed into the response information in a determined recommendation style for recommendation; a recommendation effectiveness statistics module, configured to calculate the recommendation effectiveness data of each recommendation style based on the recommendation behavior data; a style decision adjustment module, configured to determine whether the recommendation effectiveness data meets a preset condition, and when the preset condition is not met, readjust the corresponding recommendation style in the response prompt text.

[0171] Based on any embodiment of the device of the present application, the style decision adjustment module includes: a style prompt construction module, which is configured to embed each recommended style and its corresponding recommended effectiveness data into a style prompt text, wherein the style prompt text contains a style adjustment instruction, and the style adjustment instruction indicates a target recommended style that needs to be adjusted to improve its recommendation effectiveness data; a style version generation module, which is configured to input the style prompt text into a large language model to generate multiple versions of the target recommended style according to the style adjustment instruction; a style test optimization module, which is configured to input the generated multiple versions into the response business logic of the response prompt text for comparative testing, and replace the target version with the highest recommended effectiveness data in the comparative test result with the determined version in the response business logic.

[0172] Based on any embodiment of the device of the present application, the device also includes: a behavior data update module, which is configured to asynchronously update the behavior data generated by the user in the online store, and format it and store it in a behavior database, so that the behavior data in the behavior database includes a timestamp for providing time information and behavior event data for describing corresponding behavior events; a chat record update module, which is configured to asynchronously update the chat records generated by the user's conversation with the chat robot, and format it and store it in a record database, so that the chat records in the record database include a timestamp for providing time information and multiple rounds of conversations generated by the user's interaction with the chat robot; a text semantic extraction module, which is configured to vectorize the behavior data and chat records for semantic matching based on the short chat records between the user and the chat robot, so as to determine the associated historical behavior data and the historical chat records.

[0173] Another embodiment of the present application also provides an intelligent customer service answering device. Figure 4 As shown, a schematic diagram of the internal structure of an intelligent customer service answering device. The intelligent customer service answering device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. Among them, the computer-readable non-volatile readable storage medium of the intelligent customer service answering device stores an operating system, a database, and computer-readable instructions. The database may store an information sequence. When the computer-readable instructions are executed by the processor, the processor may implement an intelligent customer service answering method / network live broadcast automatic answering method.

[0174] The processor of the intelligent customer service answering device is used to provide computing and control capabilities to support the operation of the entire intelligent customer service answering device. The memory of the intelligent customer service answering device may store computer-readable instructions, which, when executed by the processor, may enable the processor to execute the intelligent customer service answering method / network live broadcast automatic answering method of the present application. The network interface of the intelligent customer service answering device is used to connect and communicate with the terminal.

[0175] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the intelligent customer service response device to which the solution of the present application is applied. The specific intelligent customer service response device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0176] In this embodiment, the processor is used to execute Figure 3The memory stores the program code and various data required to execute the above modules or submodules. The network interface is used to realize data transmission between user terminals or servers. The non-volatile readable storage medium in this embodiment stores the program code and data required to execute all modules in the intelligent customer service response device of this application, and the server can call the program code and data of the server to execute the functions of all modules.

[0177] The present application also provides a non-volatile readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the intelligent customer service response method / live broadcast automatic response method of any embodiment of the present application.

[0178] The present application also provides a computer program product, including a computer program / instruction, which implements the steps of the method described in any embodiment of the present application when executed by one or more processors.

[0179] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0180] In summary, this application not only improves the accuracy and personalization of response information, but also enhances the flexibility and effectiveness of product recommendations, brings significant commercial value and user satisfaction to online stores on e-commerce platforms, and has a high economy of scale utility.

Claims

1. An intelligent customer service response method, characterized in that: include: In response to a user question event, the latest short chat record between the chat robot and the user is obtained, wherein the user question event includes the latest question statement of the user; According to the short chat records, determining semantically compatible historical behavior data from the user's behavior data, and determining semantically compatible long chat records from the user's historical chat records; The short chat records, long chat records and historical behavior data are combined into question reference data, and the intended product that the user intends to consult and its corresponding intention strength are determined according to the question reference data; The question reference data, intended products and their intention strength are embedded in a preset answer prompt text, and according to the preset answer business logic of the answer prompt text, answer information corresponding to the intention strength and matching the latest question statement is generated, and the answer information includes the owned products determined from the online store based on the intended products.

2. The intelligent customer service response method according to claim 1, characterized in that: Determining the intended product that the user intends to consult and its corresponding intention strength according to the question reference data includes: The question reference data is embedded in the intention recognition prompt text, wherein the intention recognition prompt text includes an intention recognition control instruction for indicating to refer to the short chat record, the long chat record and the historical behavior data to recognize the user's latest intended product and its intention strength; The intention recognition prompt text is input into the large language model to control the large language model to determine the user's latest intended product and its intention strength according to the intention recognition control instruction and with reference to the time information and text semantics provided by the question reference data.

3. The intelligent customer service response method according to claim 1, characterized in that: The question reference data, intended products and their intended strength are embedded in a preset answer prompt text, and according to the preset answer business logic of the answer prompt text, answer information corresponding to the intended strength and matching the latest question statement is generated, and the answer business logic instructs the large language model to execute the following business process: Based on the intended product and the comprehensive semantic information of the question reference data, online searching for a set of related products matching the comprehensive semantic information; According to the description information of the associated products in the associated product set, the owned products stored in the product database of the online store are matched; Adapt the intention strength to determine the corresponding preset recommendation style in the response business logic, and construct response information corresponding to the latest question semantics according to the recommendation style.

4. The intelligent customer service response method according to claim 3, characterized in that: Adapting the intention strength to determine a corresponding preset recommendation style in the response business logic includes any one or more of the following steps: When the intention strength is the highest strength, determining a first recommendation style, the first recommendation style indicating that the self-owned product is directly and prominently recommended in the corresponding response information; When the intention strength is the second highest strength, determining a second recommendation style, the second recommendation style indicating adding associated recommendation information for enhancing the user's intention to purchase the owned product to the corresponding response information; When the intention strength is the lowest strength, a third recommendation style is determined, and the third recommendation style indicates providing the self-owned product in the corresponding response information and adding a manual agent dialogue interface.

5. The intelligent customer service response method according to claim 3, characterized in that: After generating the response information corresponding to the intention strength and matching the latest question statement, it includes: Obtaining the recommendation behavior data corresponding to the user's access to the own products in the online store after the products are constructed into the response information in a determined recommendation style for recommendation; Calculate the recommendation effectiveness data of each recommendation style according to the recommendation behavior data; Determine whether the recommendation effectiveness data meets a preset condition, and when the preset condition is not met, readjust the corresponding recommendation style in the response prompt text.

6. The intelligent customer service response method according to claim 5, characterized in that: Re-adjust the corresponding recommended styles in the response prompt text, including: embedding each recommended style and its corresponding recommended effectiveness data into a style prompt text, wherein the style prompt text includes a style adjustment instruction, wherein the style adjustment instruction indicates a target recommended style that needs to be adjusted to improve its recommended effectiveness data; Inputting the style prompt text into a large language model to generate multiple versions of the target recommended style according to the style adjustment instruction; The generated multiple versions are put into the response business logic of the response prompt text for comparative testing, and the target version with the highest recommended effectiveness data in the comparative test results is replaced with the determined version in the response business logic.

7. The intelligent customer service response method according to any one of claims 1 to 6, characterized in that: The method further comprises: Asynchronously updating the behavior data generated by the user in the online store, formatting the behavior data and storing it in a behavior database, so that the behavior data in the behavior database includes a timestamp for providing time information and behavior event data for describing corresponding behavior events; Asynchronously updating the chat records generated by the conversation between the user and the chat robot, and formatting and storing the chat records in a record database, so that the chat records in the record database include a timestamp for providing time information and multiple rounds of conversations generated by the interaction between the user and the chat robot; The behavior data and chat records are vectorized for semantic matching based on the short chat records between the user and the chat robot to determine the associated historical behavior data and the historical chat records.

8. An intelligent customer service response device, characterized in that: include: A question response module is configured to respond to a user question event and obtain the latest short chat record between the chat robot and the user, wherein the user question event includes the latest question statement of the user; A data matching module, configured to determine semantically matching historical behavior data from the user's behavior data based on the short-term chat records, and to determine semantically matching long-term chat records from the user's historical chat records; an intention identification module, configured to combine the short chat records, long chat records and historical behavior data into question reference data, and determine the intended product that the user intends to consult and its corresponding intention strength according to the question reference data; The response generation module is configured to embed the question reference data, intended products and their intention strength into a preset response prompt text, and generate response information corresponding to the intention strength and matching the latest question statement according to the preset response business logic of the response prompt text, wherein the response information includes the own products determined from the online store based on the intended products.

9. An intelligent customer service answering device, comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 7.

10. A non-volatile readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images