Customer service knowledge point automatic configuration method and system

Abstract

The invention discloses a customer service knowledge point automatic configuration method and system, and the method comprises the steps: monitoring whether a knowledge base is triggered or not when a buyer question generates customer service information, and determining that the corresponding buyer question is a target buyer question; obtaining commodity detail information of a session commodity corresponding to the question asked by the target buyer, and corresponding reference historical session data; inputting the question of the target buyer, the commodity detail information of the session commodity and the reference historical session data as learning contents into a self-learning agent; obtaining reply information output by the self-learning agent, and generating a new knowledge point based on the reply information and a corresponding target buyer question; and automatically updating the knowledge base based on the new knowledge points and the corresponding session commodities. The method and the system can automatically configure and improve the knowledge points in the knowledge base.

Description

Technical Field

[0001] This application relates to the field of e-commerce customer service technology, and in particular to a method and system for generating customer service information based on multimodal intent recognition, a method for maintaining an e-commerce customer service knowledge base, a method for automatically configuring customer service knowledge points, a method for generating customer service information based on dialogue guidance, and the corresponding systems for each method. Background Technology

[0002] Compared to traditional customer service systems that rely on keyword matching for semantic understanding, e-commerce customer service systems using AI models (large language models) possess stronger semantic understanding capabilities. E-commerce customer service system conversation data typically contains a large amount of multimodal information, including text, images (including pictures and videos), and voice. However, existing AI models can only perform shallow semantic fusion when processing multimodal information, remaining at the feature concatenation level. This makes it difficult for customer service systems to accurately identify buyer intent, thus reducing the accuracy of customer service information.

[0003] With the changing business and customer service scenarios in e-commerce, the knowledge base of e-commerce customer service systems is prone to knowledge gaps. When the knowledge base lacks knowledge, it becomes difficult to identify which buyer questions require configured knowledge, making it hard to upgrade and maintain the knowledge base in a timely manner. This results in customer service information not being generated or being generated inaccurately.

[0004] Furthermore, the knowledge base of e-commerce customer service systems contains millions of knowledge points, the vast majority of which are automatically generated based on clustering of historical conversation data. When a buyer asks a new question outside of historical conversations, the customer service system may be unable to generate corresponding customer service information because the knowledge base lacks the corresponding knowledge points. Since the number of knowledge points that need to be configured is enormous, it is difficult for users to configure each knowledge point individually.

[0005] During a customer service conversation, if a buyer enters the conversation directly without referring to a product card or order card, the customer service system cannot identify the product card or order card the buyer is inquiring about, resulting in an inability to effectively respond to the buyer's questions. In this situation, the customer service system typically forces the buyer to provide the corresponding card information. If the buyer does not provide the card information, the system will repeatedly prompt the buyer, leading to interruptions in the conversation or getting stuck in an ineffective conversation loop, thereby harming the user experience.

[0006] Other technical issues related to this application will be further elaborated below. The above content is only for assisting in understanding the technical solutions of this application and does not imply that all of the above content is prior art. Summary of the Invention

[0007] The primary objective of this application is to provide a method and system for generating customer service information based on multimodal intent recognition. This significantly improves the accuracy of buyer intent recognition in multimodal conversation data and reduces the response delay time for generating customer service information. Furthermore, this application provides a method and system for maintaining an e-commerce customer service knowledge base. This system can automatically identify buyer questions that require supplementary (configuration) knowledge, thereby improving the maintenance efficiency of the knowledge base and the accuracy of generated customer service information. Additionally, this application provides a method and system for automatically configuring customer service knowledge points. This system can automatically configure and improve knowledge points in the knowledge base, significantly improving the configuration efficiency of the knowledge base and customer service efficiency. Finally, this application provides a method and system for generating customer service information based on dialogue guidance. This system can automatically generate customer service information corresponding to buyer questions even when the buyer does not provide card information.

[0008] To achieve the above objectives, this application proposes a customer service information generation method based on multimodal intent recognition for use in e-commerce customer service systems. The method includes: Step K1: Obtain customer service conversation data, perform heterogeneous data classification on the conversation data; perform feature extraction on each modal information after classification, including: Step K11: Perform word segmentation on the text information and extract word-level semantic features, sentence-level logical features, and discourse intent features respectively; Step K12: Extract global features and local detail features of image information to obtain intrinsic attribute features and state attribute features; Step K13: Separate the voiceprint and speech content of the speech information, extract the emotional features of the voiceprint and the speech content respectively, and perform cross-modal feature fusion on the voiceprint emotional features and the speech content emotional features; Step K14: Add a spatiotemporal identifier to the object associated with the session data, the spatiotemporal identifier including time information and spatial information; Step K2: Perform cross-modal hierarchical alignment and knowledge graph injection operations, including: Step K21: Based on the features extracted in steps K11-K13, establish a semantic bridging matrix, perform hierarchical attention routing, and output intent labels; Step K22: Inject a knowledge graph, which includes a static knowledge base, real-time business flow, and user history profile, and output three sources of dynamic knowledge; Step K3: Construct a decision tree based on intent tags, three-source dynamic knowledge, and the spatiotemporal identifier. Generate a root node based on the intent classification result of the intent tags. Set branch conditions based on the business rules of the customer service scenario, user status, and real-time emotions. Generate leaf nodes based on the root node and the branch conditions. The leaf nodes are sets of response actions. Step K4: Based on the response actions corresponding to the leaf nodes of the decision tree, perform the corresponding text generation operation, visual tag generation operation, or speech synthesis operation to generate customer service information.

[0009] Other technical features and effects of the customer service information generation method of this application will be described in the latter part of the specification. The technical problem-solving approach and related product design scheme of this application are as follows: Existing AI models struggle to deeply identify buyer intent when processing customer service conversation data across different modalities simultaneously. For example, when a user (buyer) sends both a product image and the text description "This color is actually darker," the customer service system cannot establish a precise correlation between color attributes and textual reviews. Similarly, when a user complains about "delayed logistics," the customer service system only provides a perfunctory apology, failing to retrieve real-time logistics data and offer solutions. The primary reason for this is that general-purpose large language models are not specifically trained for e-commerce customer service conversation scenarios. Extracting multimodal information often generates more noise, making it difficult to align features extracted from different modalities across modal levels, and even leading to semantic conflicts.

[0010] If we extract more accurate semantic information by reducing noise, we need to pre-configure complex noise reduction rules, which takes a lot of time to configure. Moreover, the noise reduction rules are only effective for noise in specific situations.

[0011] The applicant discovered that by significantly increasing the richness of intent across all modalities and then dynamically constructing a decision tree, and by refining data monitoring and dialogue control, it is possible to accurately identify buyer intent encompassing all modalities and generate more precise customer service information. Therefore, the applicant proposes a customer service information generation method and e-commerce customer system (hereinafter referred to as AI customer service) based on multimodal intent recognition to address the problems faced by existing AI models in processing multimodal intent recognition, as detailed below.

[0012] refer to Figure 1-2In the e-commerce customer service system, a multimodal processing module, a semantic fusion module, a knowledge base module, a decision tree module, a multimodal output module, and a continuous learning module are set up. First, customer service conversation data (i.e., the latest buyer questions) is obtained through the multimodal processing module, and heterogeneous data classification is performed on the conversation data; then feature extraction is performed on each modal information after classification, including: (1) the text information is segmented by the text processing unit. After segmentation, in addition to extracting word-level semantic features, sentence-level logical features and text intent features are also extracted to further enrich the intent of the text information; (2) the global features and local detail features of the image information are extracted by the image processing unit to obtain inherent attribute features (including color, material, etc.) and state attribute features (including defects, traces of use, etc.); 3) Separate the voiceprint and voice content of the voice information through the voice processing unit, extract the emotional features of the voiceprint and the voice content respectively, and then perform cross-modal feature fusion on the voiceprint emotional features and the voice content emotional features to avoid automatic speech recognition (ASR) errors that lead to misjudgment of emotions; (4) Add spatiotemporal identifiers to the objects associated with the session data (such as order information, logistics information and other objects associated with the session data) through the time-series data processing unit. The spatiotemporal identifiers include time information and spatial information. Among them, the time information includes the order time, current time and logistics time, and the spatial information includes the user address, warehouse address and other information.

[0013] Secondly, cross-modal hierarchical alignment and knowledge graph injection operations are performed through the semantic fusion module, including: (1) establishing a semantic bridging matrix between features extracted from text information, image information and voice information, performing hierarchical attention routing, and outputting intent labels; (2) injecting a knowledge graph (equivalent to a total knowledge base) into the semantic fusion module through the knowledge base module. The knowledge graph includes a static knowledge base (including product parameters (product details information), policy terms, historical session data, etc.), real-time business flow (including inventory status, logistics location, etc.) and user historical profile (including the number of returns and exchanges, consumption level, etc.) to output three-source dynamic knowledge (i.e., knowledge corresponding to the static knowledge base, real-time business flow and user historical profile).

[0014] Furthermore, the decision tree module dynamically constructs a decision tree based on intent tags, three-source dynamic knowledge, and spatiotemporal identifiers. Then, it generates a root node based on the intent classification results of the intent tags, sets branch conditions based on the business rules of the customer service scenario, user status, and real-time sentiment, and generates leaf nodes based on the root node and the branch conditions. The leaf nodes are sets of response actions.

[0015] Furthermore, the continuous learning module can capture user intent in the session data, feed the captured intent back to the decision tree module, and then feed it back to the decision tree structure after reinforcement learning.

[0016] Finally, the multimodal output module, based on the response actions corresponding to the leaf nodes of the decision tree, performs corresponding text generation, visual tag generation, or speech synthesis operations to generate customer service information, thereby completing the answer to the user's question in the conversation data. In actual operation, the operations of each module in this application can call different AI models (external or self-built) for processing, and promptly output the results of the previous operation to the next stage for direct processing, thereby improving the coordination efficiency of the various modules in the AI ​​customer service system.

[0017] This customer service information generation method based on multimodal intent recognition can significantly improve the accuracy of buyer intent recognition in multimodal conversation data and generate more precise customer service information. Furthermore, it can answer open-domain questions and seemingly disjointed user inquiries in the conversation data, reducing the rate of manual customer service transfers (from 35% to 12%). In addition, it can shorten the response latency for generating customer service information, reducing it to less than 0.4 seconds. When calling the AI ​​model for customer service information processing, the various functional modules support each other, effectively reducing noise in the data processing process and minimizing the influence of AI model illusions.

[0018] Furthermore, this application also provides a method for maintaining an e-commerce customer service knowledge base, used in a customer service system that generates customer service information based on an AI model. The method includes: Step L1: In response to the AI ​​model generating customer service information based on each buyer's question, the customer service system sends a self-check instruction to the AI ​​model. The self-check instruction includes instructing the AI ​​model to determine whether the target knowledge points recalled from the knowledge base are sufficient to support the buyer's question. When they are insufficient to support the buyer's question, a detection flag is generated and returned. Step L2: Obtain the detection identifier returned by the AI ​​model, and determine that the buyer's question corresponding to the detection identifier is a preliminary buyer's question; Step L3: Screen preliminary buyer questions, perform cluster analysis on the preliminary buyer questions based on vector similarity, and generate subject clusters; Step L4: Count the number of initial buyer questions in each topic cluster, sort the topic clusters from largest to smallest, and output the top-ranked topic clusters as knowledge points to be configured; Step L5: Perform the knowledge point configuration operation on the knowledge points to be configured.

[0019] The technical problem-solving approach and related product design scheme for the knowledge base maintenance method in this application are as follows: E-commerce customer service systems typically pre-configure numerous knowledge points based on customer service dialogue (conversation) scenarios. Each knowledge point includes preset questions (i.e., preset buyer questions) and corresponding preset answers (i.e., preset seller answers). Each knowledge point may contain one or more preset questions and one or more preset answers. For example... Figure 5 As shown, Q1, Q2…Q m For pre-set questions, A1, A2...A m This is a pre-set response. After the customer service system receives a new buyer's question (Q0), the AI ​​model will retrieve the corresponding knowledge points from the knowledge base as target knowledge points based on the buyer's question. Then, the customer service information generation module will generate customer service information based on the target knowledge points to serve as the seller's response to the buyer's question.

[0020] As e-commerce business and customer service scenarios evolve, the content of buyer inquiries also changes, causing many knowledge points in the knowledge base to gradually become outdated (from new knowledge to old knowledge), resulting in a lack of knowledge in the knowledge base. When handling buyer inquiries about new products, the corresponding knowledge points may also be missing. When the knowledge base lacks knowledge, customer service information may not be generated or may be generated inaccurately.

[0021] When sellers discover that the customer service information generated by the customer service system in response to a buyer's question is inaccurate, they can only rely on experience to check the knowledge base one by one to identify which knowledge points (problematic knowledge points) are causing the inaccuracy, and then correct the knowledge points corresponding to the buyer's question. Because the knowledge base contains a vast number of knowledge points (millions in total), and most of these are automatically generated based on clustering of historical conversation data (historical customer service dialogues), the knowledge base will contain many knowledge points with similar meanings. Therefore, when investigating the cause, sellers need to: first identify all knowledge points with similar meanings, and then analyze the correlation between each similar knowledge point and the buyer's question (determining whether it is the cause of the inaccurate customer service information, whether the content of each knowledge point is outdated, whether it is related to other dialogue scenarios, etc.). Only after identifying the problematic knowledge points and determining that they need to be corrected can sellers make targeted corrections to the relevant knowledge points (improving the content of the knowledge points or adding new knowledge points).

[0022] This method of identifying problematic knowledge points before maintaining the knowledge base requires a significant investment of manpower and time to investigate these points. It places extremely high demands on the work experience and knowledge capabilities of maintenance personnel (sellers or other staff) (who need to be familiar with various e-commerce business scenarios, the accurate intent of each preset question, the completeness of each preset response, and the various dialogue scenarios that each knowledge point may be associated with). Maintenance personnel also need to constantly monitor the accuracy of the generated customer service information, resulting in extremely low maintenance efficiency for the knowledge base. Furthermore, this method makes it difficult to determine whether a buyer's question needs to be corrected (configured) for knowledge points, and may correct knowledge points for some buyer questions with low consultation frequency (that originally did not need to be corrected), resulting in a waste of knowledge base maintenance resources.

[0023] The applicant discovered that the AI ​​model (large language model) has a self-checking mechanism. If the self-checking mechanism is activated (executing self-checking instructions) when the AI ​​model generates customer service information based on each buyer's question, allowing the AI ​​model to determine whether the knowledge points retrieved from the knowledge base are sufficient to support the buyer's question and return a detection label (such as "supported," "YES," or "204"), then the customer service system can determine the corresponding buyer's question as a preliminary buyer's question based on the detection label, thus obtaining a large number of preliminary buyer's questions. At this point, if these preliminary buyer's questions are retrieved and clustered, many subject clusters can be generated. These subject clusters are then sorted by size, and the top-ranked subject clusters can be used as knowledge points to be configured and output to the customer service system's knowledge base configuration module. Then, the knowledge points to be configured can be manually corrected (configured) through the front-end interface of the knowledge base configuration module, or a self-learning AI model can be integrated into the knowledge base configuration module (as described below), and the knowledge points corresponding to the knowledge points to be configured can be automatically generated based on the learning results, thereby completing the maintenance of the knowledge base.

[0024] This AI-based self-checking mechanism for maintaining the knowledge base can automatically identify buyer questions that need supplementation, automatically discard low-frequency buyer questions with small topic clusters, and avoid the need for knowledge base maintenance and upgrades for a large number of low-frequency buyer questions. Furthermore, when identifying buyer questions requiring maintenance, it does not rely on the work experience and knowledge capabilities of maintenance personnel, nor does it require them to constantly monitor the accuracy of generated customer service information. In addition, it avoids the need to individually check numerous similar knowledge points during knowledge base maintenance, thus significantly improving maintenance efficiency. Because relevant knowledge points in the knowledge base are maintained in a timely manner, the accuracy of generated customer service information can be improved.

[0025] Furthermore, this application also provides a method for automatically configuring customer service knowledge points, the method comprising: Step G1: Monitor whether the buyer's question triggers the knowledge base when generating customer service information. If the buyer's question triggers the knowledge base but no knowledge points are obtained from the knowledge base, determine that the corresponding buyer's question is the target buyer's question. Step G2: Obtain the product details information of the product in the conversation corresponding to the question asked by the target buyer, wherein the product details information refers to the product information in the product details page; and obtain the historical conversation data corresponding to the product in the conversation, and perform noise reduction processing on the historical conversation data to obtain reference historical conversation data; Step G3: Input the target buyer's question, the product details of the conversational product, and the reference historical conversation data as learning content into the self-learning agent, instruct the self-learning agent to learn the learning content, and output the answer information corresponding to the target buyer's question; Step G4: Obtain the response information output by the self-learning agent, and generate new knowledge points based on the response information and the corresponding target buyer's question, wherein the target buyer's question is a preset question for the new knowledge point, and the response information is a preset answer for the new knowledge point; Step G5: Automatically update the knowledge base based on the new knowledge points and corresponding conversational items.

[0026] The technical problem-solving approach and related product design scheme for the knowledge base maintenance method in this application are as follows: Because the number of knowledge points that need to be configured in the knowledge base is very large, manual configuration of knowledge points is mainly applicable to a small number of knowledge points that buyers ask frequently, as it is difficult for users to configure each knowledge point individually.

[0027] The applicant discovered that for new buyer questions outside the knowledge base, the software program returns a "204" status code (indicating a successful request but no data return) when generating customer service information because the new buyer question cannot trigger the knowledge base. Conversely, if knowledge points are obtained from the knowledge base (instead of a "204" status code), it means the buyer question does not lack the corresponding knowledge points and no knowledge point configuration is needed. Therefore, if the receipt of a status code (or other detection indicator) is used to determine whether a buyer question lacks knowledge points, then the buyer questions that require knowledge point configuration (i.e., target buyer questions) can be automatically obtained.

[0028] The applicant also discovered that for new buyer questions outside the knowledge base, although there were no directly corresponding questions and answers in the historical session data, it was still possible to extract some response information from the historical session data for many new buyer questions. Therefore, after identifying the target buyer's question (equivalent to the initial buyer question in step L2), the historical session data of the corresponding product in the session can be denoised and used as reference historical session data. Then, the target buyer's question, the product details of the session product, and the reference historical session data are input as learning content into the self-learning agent (AI model or other agent), and the self-learning agent is instructed to output the corresponding response information after learning the learning content. In this way, new knowledge points can be generated based on the response information output by the self-learning agent and the corresponding target buyer question, and then the knowledge base can be automatically updated based on the new knowledge points.

[0029] This automatic configuration method for customer service knowledge points can automatically configure and improve the knowledge points in the knowledge base, enabling the customer service system to answer buyer questions that are outside the scope of the knowledge base, thus greatly improving the configuration efficiency of the knowledge base and the efficiency of customer service.

[0030] Furthermore, this application also provides a method and system for generating customer service information based on dialogue guidance, the method comprising: Step H1: Obtain the first buyer message from the first round of the conversation, and detect the content of the first buyer message. When the first buyer message contains a buyer's question but does not contain card information, start the dialogue guidance strategy and send a card information request to the buyer. The card information is a product card or an order card. When the first buyer message contains both a buyer's question and card information, generate customer service information based on the obtained card information, buyer's question, and knowledge base. Step H2: Obtain the second buyer message from the second round of the conversation, and check the content of the second buyer message. If the second buyer message contains card information, end the dialogue guidance strategy and generate customer service information based on the obtained card information, buyer questions, and knowledge base. If the second buyer message does not contain card information, match the corresponding basic conversation scenario from the conversation scenario library in the knowledge base according to the first and second buyer messages. Determine the type of card to be obtained based on the basic conversation scenario: if the basic conversation scenario belongs to the product shopping scenario, then the type of card to be obtained is determined to be a product card; if the basic conversation scenario belongs to the after-sales processing scenario, then the type of card to be obtained is determined to be an order card. Step H3: Generate a predicted card based on the type of card to be acquired, and send the predicted card to the buyer for confirmation; Step H4: Obtain the third buyer's message from the third round of conversation, and examine the content of the third buyer's message. If the third buyer's message contains card information or the buyer approves the predicted card, end the dialogue guidance strategy and generate customer service information based on the obtained card information or the approved predicted card, as well as the buyer's question and knowledge base; otherwise, transfer to human customer service and generate customer service information based on the human customer service's editing information.

[0031] The technical problem-solving approach and related product design scheme of this application based on the dialogue-guided customer service information generation method are as follows: When customer service conversations lack card information (product cards or order cards), it's difficult to generate effective customer service information (as a seller's response) based solely on the buyer's question and the customer service system's knowledge base. If the buyer is forced to provide the corresponding card information, and they consistently fail to do so, the system will get stuck in an ineffective follow-up loop. This not only fails to answer the buyer's question but also damages the buyer's customer service experience. Therefore, it's necessary to obtain the card information corresponding to the buyer's question within the fewest possible conversation rounds.

[0032] The applicant discovered that when a buyer asks for card information in the first round of the conversation, if the customer service system initially requests the corresponding card information, the buyer usually provides the card information or related explanations in the second round of the conversation. Therefore, if the buyer provides the card information correctly in the second round, the conversation can be closed, and customer service information can be generated based on the obtained card information and the buyer's question; if the buyer does not provide card information, the conversation needs to be guided further.

[0033] In related solutions, repeatedly asking buyers for card information when the second round of conversation is incomplete can easily provoke buyer resistance. Asking buyers to provide more details avoids repeating the same questions and reduces the risk of resistance. However, this requires the system to accurately determine exactly what information is missing and guide the buyer to provide supplementary answers. This necessitates pre-setting corresponding missing information judgment rules in the customer service system based on various possible business scenarios, making the system's execution logic very complex. Furthermore, when the buyer's question is not covered by the pre-set judgment rules, the system cannot know the specific missing content in the second round of conversation, making it difficult to continue guiding the dialogue.

[0034] The applicant also found that, in the absence of card information, although the corresponding card information could not be determined based on the buyer messages in the first and second rounds of conversation, the basic conversation scenario of the buyer's question could be determined, and the type of card to be obtained could be determined based on the basic conversation scenario.

[0035] Once it's determined whether the card to be retrieved is a product card or an order card, an AI model can be used to analyze the buyer's basic account information, historical purchase behavior data, store browsing behavior data, and buyer messages from the first and second rounds of conversation. This AI model then generates corresponding predicted card information. The predicted card information is then sent to the buyer for confirmation. If the buyer approves the predicted card, customer service information can be generated based on the predicted card, the buyer's question, and the customer service system's knowledge base. If the buyer does not approve the predicted card, the matter can be transferred to a human agent to answer the buyer's question.

[0036] This dialogue-guided customer service information generation method can automatically guide buyers to provide card information and automatically generate corresponding customer service information for buyer inquiries when buyers do not provide card information. It can determine the appropriate card information within three rounds of conversation, significantly reducing the number of dialogue guidance rounds and improving the buyer's customer service experience. Furthermore, it avoids the lack of judgment rules in pre-setting complex content in the customer service system and is compatible with most unknown e-commerce customer service dialogue scenarios, improving the accuracy of customer service information.

[0037] Furthermore, this application also includes systems corresponding to various methods. These systems contain the functional modules involved in this application, execute operation instructions for the corresponding functional modules or methods, and output relevant data information to the system front-end interface. The system is stored in a server and / or computer device containing a processor, which executes the system's operation instructions.

[0038] Declaration: The functional modules of this application can be integrated with each other, or exist independently, or one functional module can be a sub-module of another functional module; the methods and systems can be integrated with each other to form a new method or system; the step numbering does not limit the order of the corresponding operation steps. Attached Figure Description

[0039] The accompanying drawings are provided to further understand this application and do not constitute a limitation thereof; the content shown in the drawings may be actual data of the embodiments and falls within the protection scope of this application.

[0040] Figure 1 This is a schematic diagram illustrating the principle of a customer service information generation method based on multimodal intent recognition in one embodiment.

[0041] Figure 2 This is a schematic diagram illustrating the working principle of a multimodal processing module in one embodiment.

[0042] Figure 3 This is a schematic diagram of a customer service information generation method based on multimodal intent recognition in one embodiment.

[0043] Figure 4This is a schematic diagram of an e-commerce customer service knowledge base maintenance method in one embodiment.

[0044] Figure 5 This is a schematic diagram illustrating the principle of an e-commerce customer service knowledge base maintenance method in one embodiment.

[0045] Figure 6 This is a schematic diagram of the knowledge base configuration module structure in one embodiment.

[0046] Figure 7 This is a schematic diagram of the operation interface of the knowledge point configuration module in one embodiment.

[0047] Figure 8 This is a schematic diagram of the operation interface of the knowledge point testing module in one embodiment.

[0048] Figure 9 This is a schematic diagram of a method for automatically configuring customer service knowledge points in one embodiment.

[0049] Figure 10 This is a schematic diagram of a dialogue-guided customer service information generation method in one embodiment. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0051] refer to Figures 1-3 This application proposes a customer service information generation method based on multimodal intent recognition, including the following steps K1-K4.

[0052] Step K1: Obtain customer service conversation data, perform heterogeneous data classification on the conversation data; perform feature extraction on each modal information after classification, including: Step K11: Perform word segmentation on the text information and extract word-level semantic features, sentence-level logical features, and discourse intent features respectively; Step K12: Extract global features and local detail features of image information to obtain intrinsic attribute features and state attribute features; Step K13: Separate the voiceprint and speech content of the speech information, extract the emotional features of the voiceprint and the speech content respectively, and perform cross-modal feature fusion on the voiceprint emotional features and the speech content emotional features; Step K14: Add a spatiotemporal identifier to the object associated with the session data. The spatiotemporal identifier includes time information and spatial information.

[0053] Step K2: Perform cross-modal hierarchical alignment and knowledge graph injection operations, including: Step K21: Based on the features extracted in steps K11-K13, establish a semantic bridging matrix, perform hierarchical attention routing, and output intent labels; Step K22: Inject a knowledge graph, which includes a static knowledge base, real-time business flow, and user historical profile, and output three sources of dynamic knowledge.

[0054] Step K3: Construct a decision tree based on intent tags, three-source dynamic knowledge, and the spatiotemporal identifier. Generate a root node based on the intent classification result of the intent tags. Set branch conditions based on the business rules of the customer service scenario, user status, and real-time sentiment. Generate leaf nodes based on the root node and the branch conditions. The leaf nodes are sets of response actions.

[0055] Step K4: Based on the response actions corresponding to the leaf nodes of the decision tree, perform the corresponding text generation operation, visual tag generation operation, or speech synthesis operation to generate customer service information.

[0056] In one embodiment, the e-commerce customer service system executes steps K1, K2, K3, and K4 respectively through a multimodal processing module, a semantic fusion module, a decision tree module, and a multimodal output module. The decision tree module includes a node generation unit and a reinforcement learning unit. The node generation unit constructs a decision tree and outputs the decision tree structure. The reinforcement learning unit presets a Markov decision process and updates the policy network of the decision tree using a policy optimization algorithm (PPO algorithm).

[0057] A Markov Decision Process (MDP) comprises a state set, an action set, a transition function, and a reward function. The transition function determines the probability of a state transition after an action is performed. In a pre-defined MDP, the state set includes the current dialogue turn and the user's state; the action set includes asking for supplementary information, providing a solution, and transferring to human assistance; and the reward function's parameters include user satisfaction, problem resolution rate, and cost control level. This significantly reduces noise in processed customer service conversation data, thereby mitigating the illusionary influence of the AI ​​model. Furthermore, the MDP can be configured with conditional settings via a front-end interface, allowing users to tailor decision parameters based on store attributes to improve the learning efficiency of the reinforcement learning unit.

[0058] Furthermore, the e-commerce customer service system also includes a continuous learning module, which comprises an intent capture unit. This unit performs real-time intent capture based on intent vector confidence (OOD real-time intent capture). The AI ​​model generates candidate tags (such as "digital RMB payment consultation") based on the captured intents. After the candidate tags undergo manual review, the decision tree module is fine-tuned online based on the review results (e.g., updating relevant parameters of the decision tree). This improves the accuracy of the decision tree's decision results.

[0059] Furthermore, the continuous learning module also includes a feedback reinforcement unit, which includes a feedback reinforcement loop. The feedback reinforcement loop performs strategy updates based on explicit feedback and implicit feedback. Explicit feedback includes user satisfaction ratings (such as 1-5 stars), and implicit feedback includes the user's dwell time during the conversation and the number of follow-up questions asked in the conversation. The strategy update includes synchronizing the decision tree weights trained offline each day to the decision tree module.

[0060] In one embodiment, in step K12, it is first determined whether the image information is a product image (product picture or product video): if it is a product image, then the global features and local detail features of the image information are extracted to obtain inherent attribute features and state attribute features, wherein the inherent attribute features include the color and material of the product, and the state attribute features include defects and signs of use of the product; if it is not a product image, then the feature extraction of the image information is skipped. In this way, in return and exchange scenarios, it is possible to accurately distinguish between product defects and user damage, improving the efficiency of e-commerce customer service systems in handling return and exchange disputes.

[0061] In one embodiment, during hierarchical attention routing in step K21, entity layer, condition layer, and intent layer processing are performed sequentially. Entity layer processing includes direct matching of product name, number, and price information; condition layer processing includes logical expression processing (e.g., processing "if...then..."); and intent layer processing includes comprehensive output processing of the final intent. In this way, various hidden intents in the session data can be accurately extracted.

[0062] In one embodiment, in step K3, when the intent classification result of the intent tag is a return application, the business rules include determining whether the product is within the return period, the user status includes the user's VIP level and historical return rate, and the real-time emotion includes the user's anger index. For example, when the anger index > 0.7, an expedited operation (such as immediately transferring to manual processing) is automatically triggered.

[0063] In one embodiment, in step K4, the multimodal output module includes a script control unit. During text generation, the script control unit performs script scheduling based on preset script rules to generate corresponding text customer service information. For example, in response to a certain instruction, promotional information (such as "spend 200 and get 30 off") can be embedded through the script control unit. Thus, users can set script conditions as needed, and the e-commerce customer service system can automatically perform operations such as sending advertisements and custom messages, thereby greatly improving the ease of use and flexibility of the e-commerce customer service system.

[0064] Furthermore, in the process of generating customer service information in the customer service system, to avoid errors in user intent recognition by the AI ​​model and to reduce the impact of AI model illusion problems, this application also proposes a customer service information generation method and system based on classification scenarios. For related technical solutions, please refer to the full text of Chinese patent document CN119311840B.

[0065] In practice, the AI ​​model is first called to analyze the user intent (buyer intent) in the customer service conversation data (i.e. the buyer's question currently being asked by the user) or the corpus after the customer service conversation data is converted (the second target corpus). Based on the analysis results, the customer service conversation data is classified into preset scenarios. The preset scenarios include a first scenario, a second scenario, and a third scenario. The first scenario includes chat interaction, activity information, process consultation, and logistics inquiry. The second scenario includes product consultation. The third scenario includes after-sales processing.

[0066] (1) When the customer service conversation data belongs to the first scenario, customer service information is generated based on preset configuration information. The preset configuration information includes configuration information for responding to fixed terms or query status information, which can be provided by the AI ​​model. (2) When the customer service conversation data belongs to the second scenario, customer service information is generated based on all or part of the knowledge graph. The knowledge graph (total knowledge base) includes a static knowledge base, real-time business flow, and user historical profiles. The static knowledge base may include product parameters, policy terms, and historical conversation data, as well as target question statements (i.e., preset questions in knowledge points) and target answers (i.e., preset answers in knowledge points) generated based on historical conversation data. Specifically, the corresponding target question statements and target answer information can be generated based on historical conversation data as training corpus. The target question statements and the corresponding target answer information constitute a corpus cluster. (3) When the customer service conversation data belongs to the third scenario, customer service information is generated based on a preset information collection program. The preset information collection program is used to collect information such as unboxing videos or user complaints, and then execute the corresponding processing program (such as refund and return) or transfer to human customer service.

[0067] In the second scenario, the target question and target answer information corresponding to the user's intent can be determined from the knowledge graph or static knowledge base. Then, the AI ​​model generates customer service information based on the target question and target answer information corresponding to the user's intent. This method, based on accurate user intent identification, reduces the influence of model illusions by performing special scenario classification operations on the user intent, significantly improving the accuracy of customer service information. Furthermore, when automatically creating a knowledge base based on historical conversation data (using clustering), it avoids labeling minority language keywords in multilingual customer service scenarios.

[0068] In other embodiments, customer service information can be generated based on preset configuration information in the first scenario and based on preset information collection procedures in the third scenario. However, in the second scenario, all or part of the operation steps in the customer service information generation method based on multimodal intent recognition corresponding to steps K1-K4 are used to generate customer service information, thereby improving the accuracy of customer service information.

[0069] In another embodiment, after generating intent tags in step K2, the target question and target answer information corresponding to the intent tags can be determined from the knowledge graph or static knowledge base. Then, the AI ​​model generates customer service information based on the target question and target answer information corresponding to the intent tags.

[0070] refer to Figures 4-8 This application also proposes an e-commerce customer service knowledge base maintenance method for a customer service system that generates customer service information based on an AI model, the method including steps L1-L5.

[0071] Step L1: In response to the AI ​​model generating customer service information based on each buyer's question, the customer service system sends a self-check instruction to the AI ​​model. This self-check instruction instructs the AI ​​model to determine whether the target knowledge points retrieved from the knowledge base are sufficient to support the buyer's question. If the knowledge points are insufficient, a detection flag is generated and returned; if the knowledge points are sufficient, no detection flag is generated. The AI ​​model can be an internal model of the customer service system or an external model invoked from elsewhere.

[0072] Step L2: Obtain the detection identifier returned by the AI ​​model and determine that the buyer's question corresponding to the detection identifier is a preliminary buyer's question.

[0073] Step L3: Filter preliminary buyer questions according to set conditions (e.g., filtering at fixed intervals such as weekly or monthly, or filtering when the number of preliminary buyer questions reaches a certain value). Perform cluster analysis on the preliminary buyer questions based on vector similarity to generate subject clusters. In specific implementation, community detection algorithms can be used for cluster analysis to obtain subject clusters.

[0074] Step L4: Count the number of initial buyer questions in each topic cluster, sort the topic clusters from largest to smallest (i.e., topic clusters with larger numbers are sorted first), and output the top-ranked topic clusters (e.g., output the top 5 or 10 topic clusters) as knowledge points to be configured.

[0075] Step L5: Perform knowledge point configuration operations (improve knowledge point content or add new knowledge points) on the knowledge points to be configured through the knowledge base configuration module.

[0076] In one embodiment, when determining whether the target knowledge points recalled from the knowledge base are sufficient to support a buyer's question, the confidence level of the AI ​​model is used to determine whether it is sufficient to support the buyer's question: when the confidence level is ≥ 0.8, the target knowledge points are determined to be sufficient to support the buyer's question; when the confidence level is < 0.8, the target knowledge points are determined to be insufficient to support the buyer's question. Here, the confidence level is the "confidence value" score that the AI ​​model gives to its own generated answers, represented by a value from 0 to 1.

[0077] Furthermore, in one embodiment, when 0.3 ≤ confidence level < 0.8, the AI ​​model also returns a detection identifier and the corresponding target knowledge point according to the self-test instruction; in step L2, the detection identifier and the corresponding target knowledge point returned by the AI ​​model are obtained; in step L4, the target knowledge points corresponding to each preliminary buyer question in the topic cluster are also output. When the confidence level < 0.3, only the detection identifier can be returned, without returning the target knowledge point corresponding to the detection identifier. In this way, the AI ​​model can automatically retrieve target knowledge points that are highly similar to the preliminary buyer questions in the topic cluster, which can be used as a reference for reconfiguring knowledge points, thereby improving the efficiency of knowledge base maintenance.

[0078] In one embodiment, the generation time of each initial buyer's question is recorded (i.e., the time when the AI ​​model returns the corresponding detection identifier). The latest generation time of each initial buyer's question within a topic cluster is defined as the topic cluster time (for example, if a topic cluster contains three initial buyer questions A, B, and C, and A's generation time is the latest, then A's generation time is used as the topic cluster time). When sorting topic clusters, if multiple topic clusters are of equal size, they are sorted from latest to earliest topic cluster time, with the topic clusters with later times ranked first. This ensures that buyer questions with later times are prioritized for knowledge base maintenance.

[0079] In related embodiments, reference is made to Figure 6-8 The customer service system includes a knowledge base configuration module, which further comprises a knowledge point configuration module and a knowledge point testing module. The knowledge point configuration module performs knowledge point configuration operations, including creating, modifying, and deleting knowledge points. The knowledge point testing module tests the application effect of the configured knowledge points. During knowledge point testing, a simulated dialogue scenario is used, inputting a preset question related to the corresponding knowledge point as a buyer's question. The system then verifies whether the customer service information generated by the customer service system matches the preset response for the knowledge point. If they match, the test is successful; otherwise, the test fails.

[0080] The knowledge point configuration module includes a knowledge point filtering window, a product details display window, and a knowledge point editing window. The knowledge point filtering window is used to filter knowledge points. When filter conditions are set in the knowledge point filtering window, the product details display window automatically displays the product details information corresponding to the filtered knowledge points. When product details information is selected in the product details display window, the knowledge point editing window automatically displays the knowledge points corresponding to the selected product details information.

[0081] The knowledge point filter window allows you to set filter conditions such as store site, product number, product name, knowledge point name, buyer question (preset question), response strategy (preset answer), and knowledge point generation date (last update date of the knowledge point). The details of each filtered product are displayed in a list format, forming a product list.

[0082] Each knowledge point in the knowledge point editing window displays its name, creation date, preset question (buyer's question), and preset answer (response strategy). The knowledge point name, preset question, and preset answer are all editable, and preset questions and answers can be added or deleted. The knowledge point editing window also includes a "Add Question / Answer" control, used to create new knowledge points for the products currently displayed in the editing window. This improves the interactive experience of configuring knowledge points in the customer service system.

[0083] In one embodiment, the knowledge points to be configured output in step L4 can be output to the knowledge point configuration module after obtaining the corresponding product details information. The product details information is displayed in the product details display window, and the corresponding preliminary buyer questions are displayed in the knowledge point editing window, so that maintenance personnel can quickly perform knowledge point configuration operations.

[0084] In one embodiment, the knowledge base configuration module automatically monitors the operational status of the corresponding store's products. When a product in a corresponding store is taken off the shelves, a product removal mark is generated in the corresponding product details information displayed in the product details display window (such as marking "taken off the shelves" in the product image position), and the knowledge point corresponding to the removed product is automatically closed in the knowledge point editing window.

[0085] In one embodiment, the knowledge point configuration module further includes a knowledge setting unit. This unit is used to set the effective store sites and effective basic session scenarios, and applies the knowledge points corresponding to the current knowledge point editing window to the set store sites and basic session scenarios. This avoids the need to repeatedly configure knowledge points when the same product is sold on different store sites, and avoids the need to repeatedly configure knowledge points when the same basic session scenario exists for different products, thereby improving the efficiency of knowledge point configuration. The basic session scenario refers to a high-frequency session scenario that occurs in customer service conversations for multiple different products, such as inquiries about compatible models, device battery issues, Bluetooth usage, fast charging support, size, material, product quality, product differences, production date and shelf life, applicable users, order status, payment method, inventory status, and logistics status—session scenarios with common attribute categories across different products. In practice, dozens or hundreds of basic conversation scenarios can be preset in the customer service system to initially classify customer service conversation scenarios. This allows knowledge points to be configured according to the different attributes of each basic conversation scenario. Moreover, when handling buyer inquiries, the buyer inquiries can be matched with each basic conversation scenario to see if they can be matched. If they are matched, the scope of customer service information responses can be determined in advance, thereby improving the accuracy of the generated responses.

[0086] In one embodiment, the knowledge point configuration module further includes a custom question and answer unit, which is used to create new knowledge points and associate them with corresponding products. The knowledge points can be created manually or in batches. Manual addition creates a single knowledge point, while batch import creates knowledge points in batches based on the content of the imported template file.

[0087] In one embodiment, the knowledge point configuration module further includes a knowledge base insight unit. This unit provides multiple ways to configure knowledge points and view knowledge base information, allowing different types of users to choose the appropriate knowledge base configuration method. Specifically, a knowledge base insight entry point can be set in the knowledge point configuration module and / or the knowledge point testing module to facilitate user access to the knowledge base insight unit.

[0088] In one embodiment, the knowledge point testing module includes a test product setting window, a buyer question input window, and a test result display window. The test product setting window is used to set the product corresponding to the knowledge point to be tested (set through store options and product options). When a buyer question is entered in the buyer question input window, the test result display window automatically displays the entered buyer question and the corresponding generated customer service information. Thus, testers can determine whether the corresponding knowledge point has been successfully implemented and whether the response effect has met expectations based on the buyer question and corresponding customer service information displayed in the test result display window.

[0089] In related embodiments, a knowledge point return link can also be generated in the test result display window. For example, a message such as "Answer incorrect? Click here to modify" can be generated, which includes the knowledge point return link. When the test result does not meet expectations, the knowledge point return link can be clicked to enter the corresponding knowledge point editing interface, thereby facilitating quick correction of the knowledge point.

[0090] In one embodiment, the knowledge point testing module further includes a result feedback unit. This unit is used to identify test results, such as marking the tested knowledge point as "test passed," thus better distinguishing whether the test was completed. Specifically, the result feedback unit can be equipped with a "like" or "dislike" button. Testers or users can like or dislike the corresponding knowledge point. Clicking the "like" button indicates that the user is satisfied with the generated answer, while clicking the "dislike" button indicates that the user is dissatisfied with the generated answer.

[0091] In one embodiment, the knowledge point testing module further includes a configuration status display window, which displays in real time the number of configured products, configured questions, and answered sessions for the user's store. The number of configured products refers to the number of products in the store whose knowledge points have been configured. For example, if the user's store has a total of 50 products, and 16 products have completed knowledge point configuration, then the number of configured products can be identified as "16 / 50". If these 16 configured products contain a total of 48 knowledge points, then the number of configured questions is 48. If 203 buyer questions have been answered based on these 48 configured knowledge points, then the number of answered sessions is 203. Thus, the configuration status display window allows for real-time monitoring and recording of the number of configured products, the number of configured knowledge points, and the frequency of knowledge point usage in the store, enabling users to understand the actual application status of the knowledge base and providing data for subsequent improvements and upgrades.

[0092] refer to Figure 9 and Figures 5-8 This application also proposes an automatic configuration method for customer service knowledge points, used in a customer service system that generates customer service information based on an AI model, the method comprising steps G1-G5.

[0093] Step G1: Monitor whether the buyer's question triggers the knowledge base when generating customer service information. If the buyer's question triggers the knowledge base but no knowledge points are obtained from the knowledge base, determine that the corresponding buyer's question is the target buyer's question.

[0094] Step G2: Obtain the product details information of the product corresponding to the question asked by the target buyer. The product details information refers to the product information on the product details page (including product title, images, parameters, unit price, material, functional attributes, operation instructions, etc.); and obtain the historical session data corresponding to the product, and perform noise reduction processing on the historical session data to obtain reference historical session data. The product in the session can be determined based on the product card, order card, or other product information contained in the corresponding session message. The noise reduction processing includes filtering out automatic reply messages generated by the customer service system without going through the knowledge base (such as reassurance or apology messages automatically generated according to preset rules and preset configuration information after the system detects abnormal user emotions), automatic messages released by the e-commerce platform (such as platform announcements and notifications), reply messages generated by the AI ​​model, invalid characters, etc. Filtering out reply messages generated by the AI ​​model allows the reference historical session data to mainly retain the content of human responses, thereby improving the learning efficiency of the self-learning agent.

[0095] Step G3: Input the target buyer's question, the product details of the conversational product, and the reference historical conversation data as learning content into the self-learning agent, instruct the self-learning agent to learn the learning content, and output the response information corresponding to the target buyer's question.

[0096] Step G4: Obtain the response information output by the self-learning agent, and generate new knowledge points based on the response information and the corresponding target buyer's question, wherein the target buyer's question is a preset question for the new knowledge point, and the response information is a preset answer for the new knowledge point.

[0097] Step G5: Automatically update the knowledge base based on the new knowledge points and corresponding conversational items, and automatically save the new knowledge points to the knowledge base configuration module.

[0098] In one embodiment, the knowledge base configuration module of the customer service system is equipped with a self-learning switch. When the self-learning switch is turned on, the knowledge base configuration module initiates a round of self-learning of knowledge points at preset intervals. During each round of self-learning, the knowledge base configuration module automatically acquires questions from target buyers within the corresponding period and inputs these automatically acquired questions, along with corresponding product details and historical conversation data, as learning content into the self-learning agent for learning. For example, a round of learning can be initiated at a fixed time each week or month, with the learning time set to start at 00:00. After completing this round of knowledge point learning, the agent accepts the next round of learning, thus saving server computing resources.

[0099] In one embodiment, the knowledge base insight unit can also automatically configure knowledge points for specified products. Upon entering the knowledge base insight unit, in response to the product specified by the user, the knowledge base configuration module automatically obtains all target buyer questions corresponding to the product to be configured, and inputs these target buyer questions, along with relevant product details and historical conversation data, as learning content into the self-learning agent. This allows for the automatic configuration of knowledge points for specific products to be initiated in advance as needed, without waiting for the next preset cycle, thereby improving the convenience of knowledge point configuration.

[0100] In one embodiment, the knowledge point configuration module includes a knowledge setting unit. This unit has a product details learning control key, which controls whether product details information is used as learning content. When the product details learning control key is enabled, product details information is used as learning content (in this case, the learning content in step G3 includes the target buyer's question, product details information, and reference historical session data). When the product details learning control key is disabled, product details information is not used as learning content (in this case, the learning content in step G3 only includes the target buyer's question and reference historical session data). This allows for flexible adjustment of the self-learning agent's learning content. By disabling the product details learning control key, automatic clustering of historical session data can be achieved, thereby improving the efficiency of automatic creation / configuration of regular knowledge points.

[0101] refer to Figure 10 and Figure 5 This application also proposes a dialogue-guided customer service information generation method for a customer service system that generates customer service information based on an AI model, the method comprising steps H1-H4.

[0102] Step H1: Obtain the first buyer message from the first round of the conversation, and detect the content of the first buyer message. When the first buyer message contains a buyer's question but does not contain card information, start the dialogue guidance strategy, send a card information request to the buyer, and continue the second round of the conversation. The card information is a product card or an order card. When the first buyer message contains both a buyer's question and card information, end the dialogue guidance strategy and generate customer service information based on the obtained card information, buyer's question, and knowledge base.

[0103] Step H2: Obtain the second buyer message from the second round of the conversation, and check the content of the second buyer message. If the second buyer message contains card information, end the dialogue guidance strategy and generate customer service information based on the obtained card information, buyer questions, and knowledge base. If the second buyer message does not contain card information, match the corresponding basic conversation scenario from the conversation scenario library in the knowledge base according to the first and second buyer messages. Based on the basic conversation scenario, determine the type of card to be obtained (the card to be obtained refers to the card that the buyer has not yet provided and that corresponds to the buyer's question): if the basic conversation scenario belongs to the product shopping scenario, then the type of card to be obtained is determined to be a product card; if the basic conversation scenario belongs to the after-sales processing scenario, then the type of card to be obtained is determined to be an order card.

[0104] Step H3: Generate a predicted card based on the type of card to be acquired, and send the predicted card to the buyer for confirmation.

[0105] Step H4: Obtain the third buyer's message from the third round of conversation, and examine the content of the third buyer's message. If the third buyer's message contains card information or the buyer approves the predicted card, end the dialogue guidance strategy and generate customer service information based on the obtained card information or the approved predicted card, as well as the buyer's question and knowledge base; otherwise, transfer to human customer service and generate customer service information based on the human customer service's editing information.

[0106] In one embodiment, the basic conversation scenarios in the conversation scenario library include major scenarios and corresponding sub-categories. The major scenarios include chat interaction scenarios, product information consultation scenarios, product discount consultation scenarios, order operation consultation scenarios, logistics information consultation scenarios, after-sales service consultation scenarios, and other major scenarios. Among them, product information consultation scenarios, product discount consultation scenarios, and order operation consultation scenarios belong to product shopping scenarios, while logistics information consultation scenarios and after-sales service consultation scenarios belong to after-sales processing scenarios.

[0107] Furthermore, the subcategories for chat interaction scenarios include greeting scenarios, conversation end scenarios, and emoji scenarios. Product information inquiry scenarios include product card inquiries, product detail inquiries, product usage inquiries, and product inventory inquiries. Product discount inquiries include promotional activity inquiries, product discount inquiries, and product price negotiation inquiries. Order operation inquiries include order modification inquiries, payment issue inquiries, product warranty inquiries, order placement issue inquiries, and other order placement operation inquiries. Logistics information inquiries include delivery / receipt time inquiries, shipping location inquiries, shipping requirements inquiries, courier selection inquiries, logistics status inquiries, and non-receipt inquiries. After-sales service inquiries include order card inquiries, address modification inquiries, order cancellation inquiries, return and refund inquiries, shipping fee refund inquiries, address change inquiries, and inquiries about missing or incorrect shipments; other major scenarios include store-related inquiries, etc.

[0108] Each major and minor scenario category is associated with one or more keywords or key phrases. When a buyer message (including buyer questions and other messages sent by the buyer) contains the corresponding keywords or key phrases, the corresponding buyer message scenario is determined to match the corresponding basic conversation scenario.

[0109] In one embodiment, the knowledge base further includes preset product industry packages. Each product industry package contains knowledge points for a specific industry. Each knowledge point contains preset buyer question information and preset seller response information, and each knowledge point is associated with a corresponding basic conversation scenario. When a buyer's question matches the corresponding knowledge point and basic conversation scenario, customer service information is generated based on the preset seller response information for that knowledge point. The product industry package can be a set of highly accurate basic conversation scenarios and knowledge point data obtained by collecting historical conversation data and product detail information from the corresponding industry and training the model.

[0110] The product industry packages include one or more of the following: a general industry package compatible with all industry dialogues; a home and living industry package compatible with the furniture and living industry dialogues; an adult product industry package compatible with the adult product industry dialogues; a 3C electronics industry package compatible with the 3C electronics (a collective term for computer, communication, and consumer electronics) industry dialogues; a beauty product industry package compatible with the beauty product industry dialogues; an outdoor sports industry package compatible with the outdoor sports industry dialogues; and a maternity and baby product industry package compatible with the maternity and baby product industry dialogues.

[0111] Furthermore, step H3 includes steps H31 and H32.

[0112] Step H31: When the type of card to be obtained is a product card, acquire the buyer's basic account information, historical purchase behavior data, and store browsing behavior data as basic product data. Input the basic product data, first buyer messages, and second buyer messages as product learning content into the AI ​​model. After the AI ​​model generates a predicted product card based on the product learning content, send the predicted product card to the buyer for confirmation. For example, send the buyer a message asking, "Are you inquiring about this product card?" for further confirmation.

[0113] Step H32: When the type of card to be obtained is an order card, acquire the buyer's basic account information, historical purchase behavior data, and store browsing behavior data as the basic order data. Input the basic order data, first buyer messages, and second buyer messages as order learning content into the AI ​​model. After the AI ​​model generates a predicted order card based on the order learning content, send the predicted order card to the buyer for confirmation. For example, send the buyer a message asking, "Are you inquiring about this order card?" for further confirmation.

[0114] Furthermore, this application also provides a customer service summary generation method, which enables the customer system to automatically generate summary information of the conversation between the customer service system and the buyer when transferring the customer to human customer service, thereby facilitating the seller to quickly understand the buyer's intent when manually processing buyer messages, and thus improving customer service efficiency.

[0115] Specifically, in response to the instruction to transfer to human customer service, the system retrieves historical conversation data (historical dialogues) between the customer service system and the buyer. This historical conversation data is then denoised (e.g., system messages and invalid characters are removed). An AI model is invoked to analyze the denoised historical conversation data and generate corresponding summary information. Based on this summary information, a summary work order is created and displayed on the front-end interface of the customer service system. When the AI ​​model generates the summary information, it is instructed to generate structured summary information according to the problem description, buyer's needs, attempted solutions, buyer's emotional state, and problem priority, thereby improving the accuracy of the summary information.

[0116] The above description is merely a preferred embodiment of this application and does not limit the patent scope of this application. All equivalent modifications made based on the inventive concept of this application and the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the patent protection scope of this application.

Claims

1. A method for automatically configuring customer service knowledge points, characterized in that, A customer service system for generating customer service information based on an AI model, the method comprising: Step G1: Monitor whether the buyer's question triggers the knowledge base when generating customer service information. If the buyer's question triggers the knowledge base but no knowledge points are obtained from the knowledge base, determine that the corresponding buyer's question is the target buyer's question. Step G2: Obtain the product details information of the product in the conversation corresponding to the question asked by the target buyer, wherein the product details information refers to the product information in the product details page; and obtain the historical conversation data corresponding to the product in the conversation, and perform noise reduction processing on the historical conversation data to obtain reference historical conversation data; Step G3: Input the target buyer's question, the product details of the conversational product, and the reference historical conversation data as learning content into the self-learning agent, instruct the self-learning agent to learn the learning content, and output the answer information corresponding to the target buyer's question; Step G4: Obtain the response information output by the self-learning agent, and generate new knowledge points based on the response information and the corresponding target buyer's question, wherein the target buyer's question is a preset question for the new knowledge point, and the response information is a preset answer for the new knowledge point; Step G5: Automatically update the knowledge base based on the new knowledge points and corresponding conversational items.

2. The method as described in claim 1, characterized in that, The noise reduction process includes filtering out automated reply messages generated by the customer service system without going through the knowledge base, automated messages released by the e-commerce platform, and reply messages generated by the AI ​​model.

3. The method as described in claim 1, characterized in that, The customer service system's knowledge base configuration module is equipped with a self-learning switch. When the self-learning switch is turned on, the knowledge base configuration module starts a round of knowledge point self-learning at preset intervals. During each round of knowledge point self-learning, the knowledge base configuration module automatically obtains the target buyer's questions within the corresponding period and inputs the automatically obtained target buyer's questions, corresponding product details information, and reference historical conversation data as learning content into the self-learning agent for learning.

4. The method as described in claim 1, characterized in that, The customer service system includes a knowledge base configuration module. New knowledge points are automatically saved to the knowledge base configuration module. The knowledge base configuration module includes a knowledge point configuration module and a knowledge point testing module. The knowledge point configuration module is used to perform knowledge point configuration operations, including creating, modifying, and deleting knowledge points. The knowledge point testing module is used to test the application effect of the configured knowledge points.

5. The method as described in claim 4, characterized in that, The knowledge point configuration module includes a knowledge point filtering window, a product details display window, and a knowledge point editing window. The knowledge point filtering window is used to filter knowledge points. When filter conditions are set in the knowledge point filtering window, the product details display window automatically displays the product details information of the product corresponding to the filtered knowledge point. When a product detail is selected in the product details display window, the knowledge point editing window automatically displays the knowledge points corresponding to the selected product detail.

6. The method as described in claim 4, characterized in that, The knowledge point configuration module also includes a knowledge base insight unit. The knowledge point configuration module and / or the knowledge point testing module contain a knowledge base insight entry point, through which one can access the knowledge base insight unit.

7. The method as described in claim 6, characterized in that, Upon entering the knowledge base insight unit, in response to the specified product to be configured, the knowledge base configuration module automatically obtains all target buyer questions corresponding to the product to be configured, and inputs the target buyer questions, corresponding product details, and reference historical conversation data as learning content into the self-learning agent for learning.

8. The method as described in claim 4, characterized in that, The knowledge point configuration module includes a knowledge setting unit, which has a product details learning control key. The product details learning control key is used to control whether to use product details information as learning content: when the product details learning control key is enabled, product details information is used as learning content. When the product details learning control is turned off, product details information will not be used as learning content.

9. The method as described in claim 8, characterized in that, The knowledge setting unit is used to set the effective store site and effective basic conversation scenario, and to apply the knowledge point corresponding to the current knowledge point editing window to the set store site and the set basic conversation scenario. The basic conversation scenario refers to the high-frequency conversation scenario that appears in the customer service conversations of multiple different products.

10. An e-commerce customer service system, characterized in that, The e-commerce customer service system is used to execute the operation instructions contained in the automatic configuration method of customer service knowledge points according to any one of claims 1-9.

Images