Customer VOC intelligent response method and system based on e-commerce platform

Abstract

The invention provides a customer VOC intelligent response method and system based on an e-commerce platform, and relates to the technical field of e-commerce platform customer services. The method comprises the following steps: recording a language text input by a client and an initial response text fed back to the client by an intelligent customer service according to the language text; receiving a subsequent input text after the client receives the initial response text; according to the subsequent input text, judging whether the understanding disorder exists or not, if so, generating an understanding disorder signal, and marking the initial response text as a to-be-adjusted text; according to the understanding obstacle signal, adjusting the to-be-adjusted text, and generating an adjusted response text; and sending the adjusted response text to the corresponding client. According to the method, the problem that the elderly customer is difficult to understand the response content of the standardized intelligent customer service is solved, the understanding obstacle of the elderly customer can be effectively identified, and the response mode of the intelligent customer service is automatically adjusted, so that the communication efficiency and the user satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of customer service of an e-commerce platform, and in particular to a method and system for intelligently responding to customer VOC based on an e-commerce platform. Background Art

[0002] In order to improve user experience and operational efficiency, e-commerce platforms generally build intelligent customer service systems to receive and quickly respond to customer inquiries, suggestions or complaints. These systems usually use advanced natural language processing (NLP) technology to parse the text entered by customers, identify the customer's true intentions, extract key information, and generate corresponding replies based on a preset knowledge base or generation model and send them to customers. The effectiveness of this process is highly dependent on the system's accurate understanding of the customer's language expression and whether it can provide appropriate, timely and clear responses.

[0003] However, when the service objects of intelligent customer service are specifically elderly customers, the existing system faces significant challenges. Due to factors such as age and cognitive habits, elderly customers are often unfamiliar and confused with the professional terms or Internet buzzwords widely used in e-commerce platforms. For example, words such as "SKU", "coupon", "flash sale", "group order", and "live streaming" may not be fully understood by elderly customers who are not often exposed to the e-commerce environment. When the intelligent customer service system includes these terms in the generated response content, it is difficult for elderly customers to accurately grasp the core information conveyed by the response, resulting in interruption of information transmission.

[0004] Due to the difficulty in understanding the responses generated by the intelligent customer service, elderly customers may show specific behavior patterns during the communication process. They may repeat the same questions as before, or try to rephrase the original questions in a language that is more familiar to them and different from the standard expression.

[0005] Existing intelligent systems often lack the ability to judge the user's understanding status when processing these subsequent inputs. The system may simply identify repeated or restated questions as a new consultation intention, thereby generating a response that is similar to the previous one but still contains difficult terms, or generate a response that is not completely relevant to the customer's actual request due to misunderstanding of the restatement. This failure to identify user understanding difficulties and make adaptive adjustments makes it easy for the dialogue process to fall into the dilemma of "speaking at cross purposes", deviating from the direction of problem solving, and forming a cycle of ineffective communication.

[0006] Existing intelligent customer service systems usually respond in a standardized language style and terminology for ordinary users, failing to dynamically adjust according to the cognitive characteristics and understanding levels of the elderly customer group. The system lacks a mechanism to perceive in real-time or near real-time that the elderly customer fails to understand the previous response and accordingly change the subsequent communication strategy and content expression. This rigid response mode makes it impossible for the elderly customer to obtain a clear and understandable solution after multiple attempts to communicate with the intelligent customer service, and the problem cannot be effectively solved.

[0007] The ineffective communication over a long period not only consumes the time and energy of the elderly customers, but also makes them have negative emotions such as confusion, frustration and even helplessness. Furthermore, it seriously doubts the effectiveness, intelligence and reliability of the intelligent customer service system, significantly reducing the trust in the intelligent customer service. This decline in trust may cause the elderly customers to abandon the convenient channel of the intelligent customer service and turn to seek the help of the human customer service team, thus increasing the workload of the human customer service team. More importantly, the failure to solve the problem in a timely and effective manner will directly affect the shopping experience of the elderly customers on the e-commerce platform, possibly leading them to abandon the purchase or even lose customers, which is not conducive to the long-term development of the e-commerce platform and the expansion of the user group. Summary of the Invention

[0008] The purpose of the present invention is to provide a customer VOC intelligent response method and system based on an e-commerce platform, which overcomes the problem that the elderly customers have difficulty in understanding the standardized intelligent customer service response content, can effectively identify the understanding obstacles of the elderly customers, and automatically adjust the response mode of the intelligent customer service to improve the communication efficiency and user satisfaction.

[0009] In a first aspect, the present invention provides a customer VOC intelligent response method based on an e-commerce platform, including the following steps: S1. Record the language text input by the customer and the initial response text feedback by the intelligent customer service to the customer for the language text; S2. Receive the subsequent input text of the customer after receiving the initial response text; S3. According to the subsequent input text, judge whether there is an understanding obstacle. If so, generate an understanding obstacle signal and mark the initial response text as the text to be adjusted; if not, input the subsequent input text as the new language text into the intelligent customer service, and return to execute step S1; S4. Adjust the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; S5. Send the adjusted response text to the corresponding customer.

[0010] The customer VOC intelligent response method based on an e-commerce platform provided by the present invention analyzes the subsequent input of elderly customers after receiving the intelligent customer service response, identifies signs of possible comprehension obstacles, and adjusts the expression mode of the response content of the intelligent customer service accordingly.

[0011] Further, between step S2 and step S3, it also includes: S6. Judging the current consultation intention of the customer according to the subsequent input text, specifically including the steps: S61. Using a preset intention recognition model, by analyzing the subsequent input text, obtain the current consultation intention of the customer; the intention recognition model is based on historical elderly customer conversation data and is fine-tuned and trained using the BERT model, capable of identifying common consultation intentions of elderly customers and outputting intention confidence levels at the same time; S62. Compare the current consultation intention with the historical consultation intention to judge whether they are consistent; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence level; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence level of the current consultation intention is greater than the preset threshold and greater than the intention confidence level of the historical consultation intention, it is determined that intention drift has occurred, and the subsequent input text is used as the new language text and input into the intelligent customer service, then return to execute step S1, otherwise execute step S3.

[0012] Further, the specific steps in step S61 include: Construct a product knowledge graph containing multiple product information; the product information includes various product attributes of various products; By parsing the subsequent input text, extract all keywords in the subsequent input text, and based on the product knowledge graph, map each keyword to a corresponding product attribute vector; Input all the product attribute vectors into the intention recognition model respectively to obtain multiple intention categories and corresponding intention confidence levels; Select the intention category with the highest confidence level as the current consultation intention of the customer.

[0013] Further, the specific steps in step S3 include: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text, and construct them into a text feature vector; S32. Input the text feature vector into a preset comprehension obstacle detection model to obtain a comprehension obstacle probability; the comprehension obstacle detection model is based on historical elderly customer conversation data and is trained using the support vector machine (SVM) algorithm, and can identify common question, negative, and repetitive expressions used by elderly customers; S33. Determine whether the comprehension obstacle probability is greater than a preset probability threshold. If it is greater, it is determined that there is a comprehension obstacle, a comprehension obstacle signal is generated, and the initial response text is marked as a text to be adjusted; otherwise, it is determined that there is no comprehension obstacle, and the subsequent input text is used as a new language text and input into the intelligent customer service, and step S1 is returned for execution.

[0014] Further, the specific steps in step S31 include: Construct a stop word list; the stop word list includes common modal particles, auxiliary words, and meaningless words; Use the stop word list to clean the subsequent input text to remove the stop words in the subsequent input text, and obtain the cleaned subsequent input text; Use a word segmentation algorithm to perform word segmentation on the cleaned subsequent input text to obtain multiple text segments; For each text segment, determine whether it belongs to a preset question word library, negative word library, or repetitive word library; if it belongs to the question word library, mark the text segment as a question word; if it belongs to the negative word library, mark the text segment as a negative word; if it belongs to the repetitive word library, mark the text segment as a repetitive word; Count the number of question words, negative words, and repetitive words in the subsequent input text, and record them as the number of question words, the number of negative words, and the number of repetitive words respectively; Construct the text feature vector according to the number of question words, the number of negative words, and the number of repetitive words.

[0015] Further, the specific steps in step S4 include: S41. Determine the adjustment method for the text to be adjusted according to the comprehension obstacle signal; the adjustment methods include term simplification and sentence pattern simplification. The term simplification is used when the comprehension obstacle signal indicates that the customer has an obstacle in understanding professional terms, and the sentence pattern simplification is used when the comprehension obstacle signal indicates that the customer has an obstacle in understanding long sentences; S42. If it is determined that the adjustment direction is only term simplification, then consult the preset vocabulary library friendly to elderly customers, calculate the semantic similarity between each standard term in the text to be adjusted and the terms in the vocabulary library friendly to elderly customers, and select the terms with a semantic similarity higher than the preset value as replacement terms, and adjust the priority of the replacement terms according to the weight value, and preferentially use the replacement term with the highest weight value to generate the adjusted response text; the terms in the vocabulary library friendly to elderly customers are all set with weight values, and the weight value represents the usage frequency of the term in the elderly customer group; S43. If it is determined that the adjustment direction is only sentence pattern simplification, then according to the preset simplification rule set, split the long sentences in the text to be adjusted into multiple short sentences and adjust the word order to generate the adjusted response text; S44. If it is determined that the adjustment direction is term simplification and sentence pattern simplification, then by consulting the vocabulary library friendly to elderly customers, replace the standard terms in the text to be adjusted with simplified expressions, and by applying the simplification rule set, perform sentence pattern and structure adjustment on the text to be adjusted after replacing the terms to generate the adjusted response text.

[0016] In a second aspect, the present invention provides a customer VOC intelligent response system based on an e-commerce platform, including: A recording module, configured to record the language text input by the customer and the initial response text fed back by the intelligent customer service to the customer for the language text; A receiving module, configured to receive the subsequent input text of the customer after receiving the initial response text; A first judgment module, configured to judge whether there is an understanding obstacle according to the subsequent input text. If so, generate an understanding obstacle signal and mark the initial response text as the text to be adjusted; if not, input the subsequent input text as a new language text into the intelligent customer service and return to run the recording module; A generating module, configured to adjust the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; A sending module, configured to send the adjusted response text to the corresponding customer.

[0017] Further, it further includes a second judgment module, and the second judgment module is configured to judge the current consultation intention of the customer according to the subsequent input text, and specifically execute the following steps: S61. Utilize the preset intention recognition model to obtain the current consultation intention of the customer by analyzing the subsequent input text; the intention recognition model is based on historical dialogue data of elderly customers and is fine-tuned and trained using the BERT model, and can recognize the common consultation intentions of elderly customers and output intention confidence at the same time; S62. Compare the current consultation intention with the historical consultation intention to determine whether they are the same; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence level; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence level of the current consultation intention is greater than a preset threshold and greater than the intention confidence level of the historical consultation intention, it is determined that an intention drift has occurred. After inputting the subsequent input text as a new language text into the intelligent customer service, return to run the recording module; otherwise, run the first judgment module.

[0018] Further, when the second judgment module is used to obtain the current consultation intention of the customer by analyzing the subsequent input text using a preset intention recognition model, it performs: Construct a product knowledge graph containing multiple product information; the product information includes various product attributes of various products; By parsing the subsequent input text, extract all keywords in the subsequent input text, and based on the product knowledge graph, map each keyword to a corresponding product attribute vector; Input all the product attribute vectors into the intention recognition model to obtain multiple intention categories and corresponding intention confidence levels; Select the intention category with the highest confidence level as the current consultation intention of the customer.

[0019] Further, when the first judgment module is used to determine whether there is an understanding obstacle according to the subsequent input text, if so, generate an understanding obstacle signal and mark the initial response text as a text to be adjusted; if not, input the subsequent input text as a new language text into the intelligent customer service and return to run the recording module, it performs: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text and construct them into a text feature vector; S32. Input the text feature vector into a preset understanding obstacle detection model to obtain an understanding obstacle probability; the understanding obstacle detection model is based on historical old customer conversation data and is trained using the support vector machine (SVM) algorithm, and can identify the common interrogative, negative, and repeated expression ways of old customers; S33. Judge whether the understanding obstacle probability is greater than a preset probability threshold. If it is greater, it is judged that there is an understanding obstacle, generate an understanding obstacle signal, and mark the initial response text as a text to be adjusted; otherwise, it is judged that there is no understanding obstacle, input the subsequent input text as a new language text into the intelligent customer service, and return to run the recording module.

[0020] As can be seen from the above, the customer VOC intelligent response method based on the e-commerce platform provided by the present invention can actively and intelligently identify whether there is an understanding barrier by analyzing the subsequent conversation behavior of elderly customers after receiving the intelligent customer service response. By dynamically adjusting the response content based on the elderly customer-friendly vocabulary library and the response simplification rule set, the standardized complex expressions are transformed into simplified expressions that are easier for elderly customers to understand. This effectively avoids the ineffective repeated communication and conversation deviation caused by language barriers, improves the accuracy and efficiency of the intelligent customer service information transmission, significantly improves the experience of elderly customers using the intelligent customer service, enhances their trust in the intelligent customer service, and indirectly reduces the pressure on the human customer service.

[0021] Other features and advantages of the present invention will be described in the subsequent specification, and, in part, will be obvious from the specification, or can be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the written specification and the accompanying drawings. Brief Description of the Drawings

[0022] Figure 1 It is a flowchart of a customer VOC intelligent response method based on the e-commerce platform provided by an embodiment of the present invention.

[0023] Figure 2 It is a schematic structural diagram of a customer VOC intelligent response system provided by an embodiment of the present invention.

[0024] Label Description: 100, recording module; 200, receiving module; 300, first judgment module; 400, generating module; 500, sending module; 600, second judgment module. Detailed Embodiments

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention usually described and illustrated in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present invention claimed, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, terms such as "first" and "second" are only used for distinguishing descriptions and cannot be construed as indicating or implying relative importance.

[0027] Referring to the attached Figure 1 , the present invention provides a customer VOC intelligent response method based on an e-commerce platform, including the following steps: S1. Record the language text input by the customer and the initial response text feedback by the intelligent customer service to the customer for the language text; S2. Receive the subsequent input text of the customer after receiving the initial response text; S3. According to the subsequent input text, determine whether there is an understanding obstacle. If so, generate an understanding obstacle signal and mark the initial response text as a text to be adjusted; if not, use the subsequent input text as a new language text and input it into the intelligent customer service, and return to execute step S1; S4. Adjust the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; S5. Send the adjusted response text to the corresponding customer.

[0028] Step S1 is responsible for capturing the interaction content between the customer and the intelligent customer service, including the original question of the customer and the reply given by the intelligent customer service. These records provide the basic data for subsequent analysis.

[0029] Step S2 is responsible for obtaining the further input of the customer after seeing the reply of the intelligent customer service. This subsequent input is the key information source for the system to judge whether the customer understands the previous reply.

[0030] Step S3 is the core link of understanding obstacle detection. It analyzes the understanding state of the customer according to the subsequent input of the customer. If the subsequent input indicates that the customer fails to understand the previous reply, for example, the customer repeats the question or expresses confusion, the system will generate an understanding obstacle signal and identify the initial response text that causes the customer's confusion, preparing for modification. If the subsequent input indicates that the customer understands the previous reply or raises a new question, the system will regard this subsequent input as a new starting point of the conversation, pass it to the intelligent customer service for processing, and the conversation process returns to step S1 to start a new interaction round.

[0031] Step S4 is started after receiving the understanding obstacle signal. For the identified text to be adjusted, modify it in terms of content or form to generate an adjusted response text that is easier for the customer to understand. Step S5 sends the adjusted and clearer response text to the customer in order to solve the previous communication obstacle and enable the customer to understand and continue the conversation.

[0032] Specifically, this method forms a feedback loop by introducing a judgment link for the customer's understanding state and triggering the adjustment of the response text according to the judgment result. First, the intelligent customer service receives the customer's language text and generates an initial response text, and these interaction contents are recorded. Then, the system receives the subsequent input text from the customer after receiving the initial response. The system analyzes whether the customer has difficulty understanding the initial response based on this subsequent input text. If it is determined that there is an understanding obstacle, the system generates an understanding obstacle signal and marks the initial response text as the text to be adjusted. Then, the system modifies the text to be adjusted according to this understanding obstacle signal to generate an adjusted response text. Finally, the adjusted response text is sent to the customer. If it is determined that there is no understanding obstacle, the subsequent input text is used as the new language text and input into the intelligent customer service, and it returns to the step of recording the initial interaction, and the intelligent customer service processes the conversation normally. This mechanism enables the intelligent customer service system to dynamically adapt to the customer's understanding level. Especially when the customer fails to understand the standard response, it can timely adjust the communication strategy and provide more understandable information, thus effectively solving the problem of ineffective communication caused by the understanding obstacle of the response text and improving the communication efficiency.

[0033] In some specific embodiments, the customer inputs the language text "How to buy this mobile phone?", and the intelligent customer service feedbacks the initial response text "Please click the 'Buy Now' button on the product page, select the SKU and quantity, and then submit the order." The system records this interaction. After the customer receives the initial response, the subsequent input text "What is SKU? Where is the button?" is input. The system receives this subsequent input text. The system analyzes according to the subsequent input text and detects that the customer has questions about "SKU" and "button", determines that there is an understanding obstacle, generates an understanding obstacle signal, and marks the initial response text as the text to be adjusted. The system adjusts the text to be adjusted according to the understanding obstacle signal. For example, "Select the SKU and quantity" is adjusted to "Select the product model and purchase quantity", and "Click the 'Buy Now' button on the product page" is adjusted to "Find the four words 'Buy Now' on the product page and tap it with your finger". The adjusted response text "Please find the four words 'Buy Now' on the product page, tap it with your finger, and then select the product model and purchase quantity and submit the order." is generated. The system sends the adjusted response text to the customer.

[0034] In certain embodiments, between step S2 and step S3, it further includes: S6. Judging the current consultation intention of the customer according to the subsequent input text, specifically including the steps: S61. Using a preset intention recognition model, by analyzing the subsequent input text, obtain the current consultation intention of the customer; the intention recognition model is based on historical dialogue data of elderly customers and is fine-tuned and trained using the BERT model, and can recognize common consultation intentions of elderly customers, and at the same time output the intention confidence; S62. Compare the current consultation intention with the historical consultation intention to determine whether they are the same; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence of the current consultation intention is greater than the preset threshold and greater than the intention confidence of the historical consultation intention, it is determined that intention drift has occurred, and the subsequent input text is used as the new language text and input into the intelligent customer service, and then return to execute step S1, otherwise execute step S3.

[0035] The intention recognition model is trained to recognize common consultation intentions of elderly customers, such as querying commodity prices, understanding commodity attributes, asking about order status, etc. By learning a large amount of historical dialogue data of elderly customers, the model can understand the unique expression methods of elderly customers. The intention confidence output by the intention recognition model represents the degree of certainty of the model for the recognition result. Comparing the current consultation intention with the historical consultation intention can be achieved through string matching or intention ID comparison. The determination logic of intention drift combines whether the intentions are the same and the confidence of the current intention. When the intentions are inconsistent and the confidence of the current intention is high enough (higher than the preset threshold) and higher than the confidence of the historical intention, the system believes that the customer's communication goal has changed significantly.

[0036] Specifically, when the intelligent customer service receives the initial language text of the customer, first recognize its historical consultation intention and record it. The intelligent customer service generates an initial response text and sends it to the customer. Subsequently, the system receives the subsequent input text of the customer. Before judging whether there is an understanding obstacle in the subsequent input text, the system uses the intention recognition model to analyze the subsequent input text to obtain the current consultation intention of the customer and its confidence. The system compares the current consultation intention with the previously recorded historical consultation intention. If the comparison result shows that the intentions are inconsistent, and the confidence of the current intention meets the preset conditions (greater than the preset threshold and greater than the confidence of the historical intention), the system determines that the customer's consultation intention has drifted. At this time, the system no longer tries to judge whether the subsequent input text is an understanding obstacle to the initial response text, but regards the subsequent input text as the starting language text of a new round of consultation initiated by the customer, and inputs it into the beginning of the intelligent customer service processing flow, that is, returns to the step of recording the customer input and the initial response of the intelligent customer service. Thus, the intelligent customer service will generate a corresponding response based on the new consultation intention of the customer. If the intention Figure 1If the current intention is the same as the historical intention or the intention is inconsistent but the confidence of the current intention is not high enough to determine a clear intention drift, the system will continue to execute the step of judging whether there is an obstacle to understanding the subsequent input text. By adding intention drift detection before judging the understanding obstacle, the system can timely identify the change of the customer's communication goal, avoid ineffective communication based on the old intention, and improve the accuracy of the intelligent customer service response.

[0037] In some specific embodiments, the customer inputs the initial language text "How to use this rice cooker?", and the intelligent customer service recognizes the historical intention as "Query of usage method" with a confidence of 0.9, and sends the initial response text, which includes the usage steps. Subsequently, the customer inputs the subsequent input text "What is the price?" The system receives this subsequent input text. The intention recognition model analyzes "What is the price?" and recognizes the current intention as "Price query" with a confidence of 0.95. The system compares the current intention "Price query" with the historical intention "Query of usage method", and the two are inconsistent. The confidence of the current intention 0.95 is greater than the preset threshold 0.7 and greater than the confidence of the historical intention 0.9. The system determines that an intention drift has occurred. Thus, the system takes "What is the price?" as the new language text and inputs it into the intelligent customer service, and returns to the recording step, and the intelligent customer service will generate a new response for the "Price query" intention.

[0038] In certain embodiments, the specific steps in step S61 include: Construct a commodity knowledge graph containing multiple commodity information; the commodity information includes various commodity attributes of various commodities, and the commodity attributes are, for example, category, brand, price range, and applicable population; By parsing the subsequent input text, extract all keywords in the subsequent input text, and based on the commodity knowledge graph, map each keyword to the corresponding commodity attribute vector; Input all commodity attribute vectors into the intention recognition model respectively to obtain multiple intention categories and the corresponding intention confidences; Select the intention category with the highest confidence as the customer's current consultation intention.

[0039] The construction of a product knowledge graph can be implemented using graph database technology, where nodes represent products, attribute types, and attribute values, and edges represent the association between products and attribute types, and between attribute types and attribute values. Extracting keywords by parsing the subsequent input text can be implemented using word segmentation technology combined with part-of-speech tagging and named entity recognition technology. Mapping keywords to product attribute vectors can be vectorized based on the knowledge graph. For example, using knowledge graph embedding technology, the entities and relationships in the knowledge graph are mapped to a low-dimensional vector space, so that semantically related entities and relationships are close in distance in the vector space. As a result, the entities or attributes corresponding to the extracted keywords can find corresponding representations in the vector space. The product attribute vector is input into the intent recognition model, which can be a classification model that receives the vector input and outputs the intent category and its confidence. Select the intent category with the highest confidence, that is, select the intent with the largest probability value output by the model as the final recognition result.

[0040] Specifically, in order to solve the problem of inaccurate intent recognition caused by the unfamiliarity of elderly customers with the products on the e-commerce platform, this solution first builds a knowledge graph containing rich product attribute information. When receiving the subsequent input text from the customer, the system parses the text and extracts the key information words. These words may not be standard product names or attribute descriptions, but by comparing and mapping with the product knowledge graph, these non-standard words can be converted into standardized product attribute representations that the system can understand, that is, product attribute vectors. For example, when a customer asks "the machine that can watch TV", the system extracts "watch TV" and "machine" as keywords, and through the knowledge graph, associates "watch TV" with "video playback" under the "function" attribute, and associates "machine" with "TV" under the "category" attribute. These associated information is converted into vector representations. Subsequently, these product attribute vectors are input into the intent recognition model trained with the conversation data of elderly customers. Based on these vectors, the model analyzes the customer's true intention, such as whether to inquire about price, function or purchase method. The model outputs multiple possible intent categories and the confidence of each category. The system selects the intent with the highest confidence as the customer's current consultation intention. In this way, even if the language used by elderly customers is not precise enough or contains daily vocabulary they are familiar with, the system can, with the assistance of the product knowledge graph, more accurately understand their consulting intentions about the product, thereby improving the accuracy of intent recognition.

[0041] In some specific embodiments, when constructing a product knowledge graph, it may include attributes such as product ID, product name, category (such as home appliances, clothing), brand (such as Haier, Midea), price range (such as 0 - 100 yuan, 100 - 500 yuan), applicable population (such as the elderly, youth), functions (such as voice control, video playback), materials, etc. For example, for a smart TV, in the knowledge graph, its category can be recorded as "TV set", brand as "a certain brand", functions include "video playback", "smart connection", "voice control", and the applicable population includes "household users". When an elderly customer inputs "How much is the machine that can be controlled by speaking with the mouth and is used for watching TV", the system extracts the keywords "speaking with the mouth to control", "watching TV", "machine", "how much". Based on the knowledge graph, "speaking with the mouth to control" is mapped to "function: voice control", "watching TV" is mapped to "function: video playback", "machine" is mapped to "category: TV set", and "how much" is mapped to "intent type: price query". These mapping results are transformed into a product attribute vector, such as a vector containing dimension information such as category, function, intent type, etc. This vector is input into the intent recognition model. After analyzing the vector, the model outputs the intent category and its confidence level. For example, the confidence level of "querying the price of the TV set" is 0.95, and the confidence level of "querying the functions of the TV set" is 0.03. The system selects the "querying the price of the TV set" with the highest confidence level as the current consultation intent of the customer. Thus, the system can accurately understand the customer's intent even if the customer uses non-standard or colloquial expressions.

[0042] In certain embodiments, the specific steps in step S3 include: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text, and construct them into a text feature vector; S32. Input the text feature vector into a preset understanding obstacle detection model to obtain the understanding obstacle probability; the understanding obstacle detection model is based on historical dialogue data of elderly customers and is trained using the support vector machine (SVM) algorithm, and can recognize common interrogative, negative, and repeated expressions of elderly customers; S33. Determine whether the understanding obstacle probability is greater than a preset probability threshold. If it is greater, it is determined that there is an understanding obstacle, an understanding obstacle signal is generated, and the initial response text is marked as a text to be adjusted; otherwise, it is determined that there is no understanding obstacle, the subsequent input text is used as a new language text and input into the intelligent customer service, and step S1 is returned for execution.

[0043] In this embodiment, by analyzing the subsequent input text of the customer, specific language features that may express confusion or lack of understanding are identified. First, in step S31, the system extracts interrogative words, negative words, and repetitive words from the subsequent input text. These types of words are often used by customers to express questions, negate the foregoing information, or emphasize parts that are not understood. Extracting these words and constructing the occurrence situations thereof into a text feature vector provides a quantitative basis for subsequent judgments. Thereby, non-standard or specific expression ways that elderly customers may use can be captured. Next, in step S32, the constructed text feature vector is input into a preset understanding obstacle detection model. This model is specifically trained based on the dialogue data of historical elderly customers and adopts the support vector machine (SVM) algorithm. The training of the model enables it to recognize the common ways of elderly customers when expressing questions, negations, or repetitions. After receiving the feature vector, the model outputs an understanding obstacle probability, which represents the possibility that the customer has an understanding obstacle. By using a model trained with data of elderly customers, the ability to recognize the unique expression ways of elderly customers is improved. Finally, in step S33, the system compares the understanding obstacle probability obtained in step S32 with a preset probability threshold. If the calculated probability is greater than the threshold, the system determines that the customer has an understanding obstacle and generates an understanding obstacle signal. At the same time, the initial response text sent by the intelligent customer service before is marked as a text to be adjusted and prepared for modification. If the probability is not greater than the threshold, it is determined that the customer does not have an understanding obstacle. The system regards the subsequent input text of the customer as a new valid input, inputs it as a new language text into the intelligent customer service, and returns to step S1 to start a new interaction round. By making a judgment based on the probability threshold, this technical solution provides a specific, operable, and highly targeted understanding obstacle detection mechanism, improves the system's ability to recognize customer confusion, and thus can trigger the adjustment of the response text more timely, improve the communication effect, and effectively enhance the accuracy and robustness of understanding obstacle recognition.

[0044] In some specific embodiments, determining whether there is an understanding barrier can be achieved as follows: After receiving the subsequent input text from the customer, the system first performs text cleaning and word segmentation. For example, a stop word list is used to remove common modal particles and meaningless words. Then, part-of-speech tagging and lexicon matching are performed on the word segmentation results. For example, a question word lexicon (such as "ma", "ne", "what"), a negation word lexicon (such as "bu", "mei", "not"), and a repeated word lexicon (such as consecutive identical words) are maintained. The number of words in the subsequent input text belonging to these three lexicons is counted to obtain the number of question words, the number of negation words, and the number of repeated words respectively. These three numbers are constructed into a three-dimensional text feature vector, such as [number of question words, number of negation words, number of repeated words]. This feature vector is input into an SVM model that has been pre-trained using elderly customer dialogue data. The model outputs a probability value between 0 and 1, indicating the likelihood of an understanding barrier. This probability value is compared with a preset probability threshold (such as 0.7). If the probability value is greater than 0.7, the system determines that the customer has an understanding barrier, generates an understanding barrier signal, and marks the response text previously sent by the intelligent customer service as pending adjustment. If the probability value is less than or equal to 0.7, it is determined that there is no understanding barrier, and the subsequent input text is processed as a new input.

[0045] The following is an example. Suppose an elderly customer consults product details on an e-commerce platform, and the intelligent customer service replies with a text containing terms such as "parameters" and "specifications". After receiving the reply, the customer inputs "I still don't understand. What does that parameter mean?".

[0046] The system receives the customer's subsequent input "I still don't understand. What does that parameter mean?", and extracts the question words "don't understand", "what does it mean", and the repeated word "still" from it. These words are converted into a numerical vector, for example, using the bag-of-words model or TF-IDF, or a more complex word embedding vector.

[0047] This vector is input into the pre-trained SVM understanding barrier detection model. This model is trained on a large amount of dialogue data of elderly customers and has learned the patterns of elderly customers expressing understanding barriers such as "didn't understand", "not clear", "say it again". After processing the vector, the model outputs an understanding barrier probability, such as 0.9.

[0048] The system compares the probability 0.9 with a preset probability threshold (e.g., 0.7). Since 0.9 > 0.7, the system determines that the customer has a comprehension barrier. At this time, the system generates a comprehension barrier signal and marks the response text containing "parameters" and "specifications" sent by the intelligent customer service before as the text to be adjusted. If the customer inputs "Okay, thank you", the model may output a low probability (e.g., 0.1), which is lower than the threshold, and the system will not generate a comprehension barrier signal nor adjust the response.

[0049] In this way, even if the expressions of elderly customers are not standard or complete enough, the system can accurately identify their comprehension barriers through key feature extraction and probability-based model judgment, triggering the subsequent response adjustment process. The specific implementation methods can include: using a natural language processing toolkit for part-of-speech tagging and keyword extraction; using pre-trained models such as Word2Vec or BERT to generate word vectors or sentence vectors; using libraries such as Scikit-learn to implement an SVM classifier and training it with the labeled dialogue data of elderly customers.

[0050] In some embodiments, the specific steps in step S31 include: Construct a stop word list; the stop word list includes common modal particles, auxiliary words, and meaningless words; Use the stop word list to clean the subsequent input text to remove the stop words in the subsequent input text, obtaining the cleaned subsequent input text; Use a word segmentation algorithm to perform word segmentation on the cleaned subsequent input text, obtaining multiple text segments; For each text segment, determine whether it belongs to a preset question word library, negative word library, or repeated word library; if it belongs to the question word library, mark the text segment as a question word; if it belongs to the negative word library, mark the text segment as a negative word; if it belongs to the repeated word library, mark the text segment as a repeated word; Count the number of question words, negative words, and repeated words in the subsequent input text, and record them as the number of question words, the number of negative words, and the number of repeated words respectively; Construct a text feature vector based on the number of question words, negative words, and repeated words.

[0051] A stop word list is constructed, which contains common modal particles, auxiliary words, and meaningless words, to reduce the interference information in the text. Subsequently, the subsequent input text is cleaned by removing the stop words to obtain a more refined text. A word segmentation algorithm is applied to decompose the cleaned text into independent lexical units, providing a basis for lexical-level analysis. Preset question word libraries, negation word libraries, and repetition word libraries are used to provide criteria for identifying specific words. Each text segmentation is judged one by one and marked according to whether it belongs to these libraries. The quantities of question words, negation words, and repetition words are counted, converting the text information into numerical features. A text feature vector is constructed based on the counted quantities, providing input data for the subsequent understanding obstacle detection model. Thus, the accuracy of extracting question words, negation words, and repetition words is improved, and the robustness of understanding obstacle recognition is enhanced.

[0052] Specifically, when receiving the subsequent input text from the customer, first, the preset stop word list is used to clean the text, removing the words that are useless for understanding obstacle judgment, such as modal particles or auxiliary words. The cleaned text is then processed by the word segmentation algorithm and decomposed into a series of independent words or phrases. Next, the system traverses these text segmentations and compares each segmentation with the preset question word libraries, negation word libraries, and repetition word libraries. If a segmentation appears in the question word library, it is marked as a question word; if it appears in the negation word library, it is marked as a negation word; if it appears in the repetition word library, it is marked as a repetition word. After all the segmentations are marked, the system counts the total number of words marked as question words, negation words, and repetition words in the subsequent input text. Finally, based on the counted quantities of question words, negation words, and repetition words, a numerical text feature vector is constructed. This text feature vector is then input into the understanding obstacle detection model to determine whether the customer has an understanding obstacle. In this way, it is possible to more accurately extract the key features reflecting understanding difficulties from the unique language expressions of elderly customers, thereby improving the effectiveness of understanding obstacle detection.

[0053] In some specific implementations, it is assumed that the subsequent input text of the customer is "Why can't I understand it? Is there no such thing?" First, a stop word list is constructed, for example, including "呢", "就是", "啊", etc. The input text is cleaned using the stop word list to obtain the cleaned text "Why can't I understand it? Is there no such thing?" Next, the cleaned text is segmented using a word segmentation algorithm to obtain a text segmentation list, for example, ["我", "怎", "弄", "不", "懂", "沒"]. Then, a preset question word library (for example, including "怎", "如何") and a negative word library (for example, including "不", "沒"). In the text segmentation list, "怎" belongs to the question word library and is marked as a question word; "不" belongs to the negative word library and is marked as a negative word; "沒" belongs to the negative word library and is marked as a negative word. The statistical results are: the number of question words is 1, the number of negative words is 2, and the number of repeated words is 0. Finally, a text feature vector is constructed based on these numbers, for example, [1, 2, 0]. The vector is input into the comprehension barrier detection model to calculate the comprehension barrier probability. By removing stop words and using a targeted vocabulary, the intentions of elderly customers to express questions and negations can be captured more accurately, improving the accuracy of feature extraction.

[0054] In some embodiments, the specific steps in step S4 include: S41. Determine an adjustment method for the text to be adjusted according to the comprehension barrier signal; the adjustment method includes terminology simplification and sentence simplification, wherein the terminology simplification is used when the comprehension barrier signal indicates that the customer has difficulty understanding professional terms, and the sentence simplification is used when the comprehension barrier signal indicates that the customer has difficulty understanding long sentences; S42. If it is determined that the adjustment direction is only term simplification, then consult the preset elderly customer-friendly vocabulary library, and calculate the semantic similarity between each standard term in the text to be adjusted and the vocabulary in the elderly customer-friendly vocabulary library, and select the vocabulary with a semantic similarity higher than the preset value as the replacement vocabulary, and adjust the priority of the replacement vocabulary according to the weight value, and give priority to the replacement vocabulary with the highest weight value to generate the adjusted response text; the vocabulary in the elderly customer-friendly vocabulary library is set with a weight value, and the weight value represents the frequency of use of the vocabulary in the elderly customer group; S43. If it is determined that the adjustment direction is only sentence simplification, then according to the preset simplification rule set, the long sentence in the text to be adjusted is split into multiple short sentences and the word order is adjusted to generate an adjusted response text so that it conforms to the reading habits of elderly customers, deletes unnecessary modifiers, and ensures that the sentence meaning is clear; S44. If it is determined that the adjustment direction is term simplification and sentence pattern simplification, then by referring to the elderly customer-friendly vocabulary library, the standard terms in the text to be adjusted are replaced with simplified expressions, and by applying the set of simplification rules, the sentence pattern and structure of the text to be adjusted after replacing the terms are adjusted to generate the adjusted response text.

[0055] Based on the received understanding obstacle signal, the system determines the way to adjust the initial response text. The understanding obstacle signal can indicate that the customer has difficulty understanding professional terms, has difficulty understanding long sentences, or has both difficulties. The system selects the corresponding adjustment strategy according to the signal type: term simplification, sentence pattern simplification, or a combination of both. When term simplification is required, the system accesses a pre-constructed elderly customer-friendly vocabulary library. This vocabulary library stores the words that are more easily understood by elderly customers and their usage frequency weights in the elderly customer group. The system calculates the semantic similarity between the professional terms in the text to be adjusted and the words in the vocabulary library, and selects the words with a similarity reaching a certain threshold as potential replacement words. Then, the system sorts these replacement words according to their weight values and preferentially selects the words with the highest weight for replacement. When sentence pattern simplification is required, the system applies a set of preset simplification rules. These rules guide the system to identify long sentences in the text to be adjusted, break them down into multiple short sentences, and adjust the arrangement order of the sentences to suit the reading habits of elderly customers. At the same time, the rules also guide the system to remove the non-core modifying components in the sentences to ensure the clarity of the core meaning of the sentences. When both term and sentence pattern simplification are required, the system first performs term replacement and then applies the sentence pattern simplification rules to the replaced text.

[0056] Specifically, when the intelligent customer service system receives a signal indicating that the customer has a comprehension obstacle, this signal triggers a process of adjusting the previously sent initial response text. First, based on the specific type of the comprehension obstacle signal, the system determines whether the customer has difficulty understanding professional terms, long sentences, or both. For example, if the comprehension obstacle signal comes from repeated inquiries about a specific word, the system may determine that term simplification is needed; if the signal comes from a confused expression about an entire sentence or paragraph, the system may determine that sentence pattern simplification is needed. After determining the adjustment method, if term simplification is required, the system will query a friendly vocabulary library designed specifically for elderly customers. This vocabulary library contains alternative words that are more familiar to elderly customers and correspond to commonly used professional terms in the e-commerce field, and each alternative word is associated with a weight value reflecting its popularity among elderly customers. The system calculates the semantic similarity between the professional terms in the text to be adjusted and the alternative words in the vocabulary library, and selects the words with high similarity as replacement candidates. To ensure the replacement effect, the system will preferentially select the alternative word with the highest weight value for replacement, thus transforming the difficult terms into easily understandable expressions. Thereby, the comprehension threshold at the term level is effectively reduced. If sentence pattern simplification is required, the system applies a set of preset simplification rule sets. These rules define how to identify and process complex sentence patterns, such as splitting long sentences containing multiple clauses into independent simple sentences, adjusting the order of sentence components to make it more in line with the reading habits of elderly customers, and deleting modifiers or insertions that do not affect the core meaning. By applying these rules, the structure and expression of the text to be adjusted are optimized, becoming more concise and clear, and reducing the comprehension difficulty brought by the complexity of sentence patterns. If the comprehension obstacle signal indicates that the customer has difficulty understanding both terms and sentence patterns, the system will first perform term replacement, replacing the professional terms in the text with friendly words, and then, on this basis, apply the sentence pattern simplification rules to adjust the structure and expression of the replaced text. This step-by-step processing method ensures that the finally generated response text has neither difficult terms nor complex sentence patterns, maximizing the comprehension efficiency of elderly customers. Through the above adjustment process, the system can generate an adjusted response text that is more in line with the cognitive characteristics of elderly customers, effectively solving the communication obstacle caused by language complexity and improving the interaction effect between the intelligent customer service and elderly customers.

[0057] In some specific embodiments, assume that the initial response text sent by the intelligent customer service to an elderly customer is "The SKU of the product you purchased has been warehoused and is expected to be shipped tomorrow. Please note to check your coupon information." The system detects that the customer has difficulty understanding "SKU" and "coupon", and determines that term simplification and sentence pattern simplification are required. The system first performs term simplification. By referring to the elderly customer-friendly vocabulary library, it is found that the friendly vocabulary corresponding to "SKU" has "product number" with a weight of 0.9 and "article number" with a weight of 0.7; the friendly vocabulary corresponding to "coupon" has "discount coupon" with a weight of 0.95 and "rebate coupon" with a weight of 0.8. The system calculates the semantic similarity and confirms that "product number" and "discount coupon" have a high similarity with the original terms. According to the weight values, the system preferentially selects "product number" to replace "SKU" and "discount coupon" to replace "coupon". The replaced text becomes "The product number you purchased has been warehoused and is expected to be shipped tomorrow. Please note to check your discount coupon information." Then, the system applies the sentence pattern simplification rule. It recognizes that this is a long sentence containing multiple pieces of information, splits it into short sentences, and adjusts the word order. For example, the rule can split the original sentence into: "The product number you purchased has been warehoused. It is expected to be shipped tomorrow. Please note to check your discount coupon information." At the same time, delete the words in "Please note to check" that may cause a comprehension burden on the elderly customer and adjust to a more direct expression. The finally generated adjusted response text is "The product number you bought has arrived at the warehouse. It is estimated that it will be shipped to you tomorrow. Please take a look at your discount coupon information." This adjusted text uses vocabulary that is more familiar to elderly customers and a simpler sentence pattern structure, significantly improving the understanding level of elderly customers.

[0058] In some specific embodiments, assume that when an elderly customer is shopping on an e-commerce platform and sees a "full reduction" activity on the product details page but does not understand its meaning, so he asks the intelligent customer service: "What does this full reduction mean?" Initial response from the intelligent customer service: "Hello, full reduction means that after the order amount reaches a certain amount, the system will automatically deduct the corresponding preferential amount. For example, full 100 minus 10 means that when the total amount of your order reaches 100 yuan, you can enjoy a 10-yuan discount." After receiving this response, the elderly customer may be confused by terms such as "order amount" and "preferential amount", or feel that the explanation is not intuitive enough, so the subsequent input is: "What is the order amount? How is the money reduction calculated?" After the system receives the subsequent input, after analysis by the understanding obstacle detection module, it generates an understanding obstacle signal, indicating that the customer has an understanding obstacle for terms such as "order amount" and "preferential amount", and indicating that the initial response text is the text to be adjusted.

[0059] Step S4 is executed: 1. According to the understanding obstacle signal, determine that the adjustment direction is term simplification.

[0060] 2. Consult the elderly customer-friendly vocabulary library and find that the simplified expressions corresponding to "order amount" are "total price of shopping" and "money you have to pay in total", and the one with a higher weight value is "total price of shopping". The simplified expressions corresponding to "discount amount" are "money saved" and "money deducted", and the one with a higher weight value is "money saved". The system preferentially selects the simplified expression with a higher weight.

[0061] 3. Apply the response simplification rule set, which may include rules such as "replace terms" and "simplify sentence patterns". The system replaces "order amount" in the original response with "total price of shopping" and "discount amount" with "money saved".

[0062] 4. Generate the adjusted response text: "Hello, the full reduction you asked about means that when the total price of shopping reaches a certain amount, the system will automatically deduct the money saved. For example, when you spend 100 yuan and save 10 yuan, it means that when the total price of your shopping reaches 100 yuan, you can save 10 yuan." The intelligent customer service sends the adjusted response to the customer. The elderly customer receives the new response, finds that easier-to-understand words are used, and successfully understands the concept of full reduction.

[0063] The specific implementation method of the elderly customer-friendly vocabulary library can be a database or a file that stores standard terms and their corresponding multiple simplified expressions and weight values. For example: "order amount": [{"expression": "total price of shopping", "weight": 0.8}, {"expression": "money you have to pay in total", "weight": 0.6}], "discount amount": [{"expression": "money saved", "weight": 0.9}, {"expression": "money deducted", "weight": 0.7}], The response simplification rule set can be a series of text conversion methods based on rules or models. For example, rule-based methods can define rules such as "delete the content in parentheses" and "convert a definition sentence starting with 'is' into an explanatory sentence starting with 'in simple terms'". Model-based methods can train a sequence-to-sequence model that takes complex text as input and generates simplified text as output.

[0064] The determination of the adjustment direction can be based on the specific content of the understanding obstacle signal. For example, if the signal contains "term not understood", then term simplification is triggered; if it contains "too long", then sentence pattern simplification is triggered. If both are included, both simplification methods are triggered simultaneously.

[0065] Refer to the appendix Figure 2 For this, the present invention provides a customer VOC intelligent response system based on an e-commerce platform, including: A recording module 100 for recording the language text input by the customer and the initial response text feedback by the intelligent customer service to the customer for the language text; A receiving module 200 for receiving the subsequent input text of the customer after receiving the initial response text; A first judgment module 300 for judging whether there is an understanding obstacle according to the subsequent input text. If so, generating an understanding obstacle signal and marking the initial response text as a text to be adjusted; if not, taking the subsequent input text as a new language text and inputting it into the intelligent customer service, and then returning to the operation recording module; A generating module 400 for adjusting the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; A sending module 500 for sending the adjusted response text to the corresponding customer.

[0066] In some embodiments, it further includes a second judgment module 600. The second judgment module 600 is used to judge the current consultation intention of the customer according to the subsequent input text, and specifically execute the following steps: S61. Using a preset intention recognition model, by analyzing the subsequent input text, obtaining the current consultation intention of the customer; the intention recognition model is based on historical dialogue data of elderly customers and is fine-tuned and trained using the BERT model, and can recognize the common consultation intentions of elderly customers and output intention confidence levels at the same time; S62. Comparing the current consultation intention with the historical consultation intention to judge whether they are consistent; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence level; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence level of the current consultation intention is greater than the preset threshold and greater than the intention confidence level of the historical consultation intention, it is determined that intention drift has occurred. After taking the subsequent input text as a new language text and inputting it into the intelligent customer service, return to the operation recording module, otherwise run the first judgment module.

[0067] In some embodiments, when the second judgment module 600 is used to obtain the current consultation intention of the customer by using a preset intention recognition model and analyzing the subsequent input text, it executes: Construct a commodity knowledge graph containing multiple commodity information; the commodity information includes various commodity attributes of various commodities, and the commodity attributes are, for example, category, brand, price range, and applicable population; By parsing the subsequent input text, extracting all keywords in the subsequent input text, and based on the commodity knowledge graph, mapping each keyword to a corresponding commodity attribute vector; Inputting all commodity attribute vectors into the intention recognition model respectively to obtain multiple intention categories and corresponding intention confidence levels; Select the intent category with the highest confidence as the current consultation intent of the customer.

[0068] In some embodiments, when the first judgment module 300 is used to judge whether there is an understanding obstacle according to the subsequent input text, if so, generate an understanding obstacle signal and mark the initial response text as the text to be adjusted; if not, use the subsequent input text as a new language text to input to the intelligent customer service, and when returning to the operation record module, execute: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text, and construct them into a text feature vector; S32. Input the text feature vector into a preset understanding obstacle detection model to obtain an understanding obstacle probability; the understanding obstacle detection model is based on historical dialogue data of elderly customers and is trained using the support vector machine (SVM) algorithm, and can identify the common interrogative, negative, and repeated expressions of elderly customers; S33. Judge whether the understanding obstacle probability is greater than a preset probability threshold. If it is greater, judge that there is an understanding obstacle, generate an understanding obstacle signal, and mark the initial response text as the text to be adjusted; otherwise, judge that there is no understanding obstacle, use the subsequent input text as a new language text to input to the intelligent customer service, and return to the operation record module.

[0069] In some embodiments, when the first judgment module 300 is used to extract the interrogative words, negative words, and repeated words in the subsequent input text and construct them into a text feature vector, it executes: Construct a stop word list; the stop word list includes common modal particles, auxiliary words, and meaningless words; Use the stop word list to clean the subsequent input text to remove the stop words in the subsequent input text and obtain the cleaned subsequent input text; Perform word segmentation on the cleaned subsequent input text using a word segmentation algorithm to obtain multiple text segments; For each text segment, judge whether it belongs to a preset interrogative word library, negative word library, or repeated word library; if it belongs to the interrogative word library, mark the text segment as an interrogative word; if it belongs to the negative word library, mark the text segment as a negative word; if it belongs to the repeated word library, mark the text segment as a repeated word; Count the number of interrogative words, negative words, and repeated words in the subsequent input text, and record them as the number of interrogative words, the number of negative words, and the number of repeated words respectively; Construct a text feature vector according to the number of interrogative words, the number of negative words, and the number of repeated words.

[0070] In some embodiments, when the generation module 400 is used to adjust the text to be adjusted according to the understanding obstacle signal and generate an adjusted response text, it executes: S41. Determine the adjustment method for the text to be adjusted according to the understanding obstacle signal; the adjustment methods include term simplification and sentence pattern simplification. Term simplification is used when the understanding obstacle signal indicates that the customer has an obstacle in understanding professional terms, and sentence pattern simplification is used when the understanding obstacle signal indicates that the customer has an obstacle in understanding long sentences. S42. If it is determined that the adjustment direction is only term simplification, then consult the preset friendly vocabulary library for elderly customers, calculate the semantic similarity between each standard term in the text to be adjusted and the terms in the elderly customer friendly vocabulary library, and select the terms with a semantic similarity higher than the preset value as replacement terms, and adjust the priority of the replacement terms according to the weight value, and preferentially use the replacement term with the highest weight value to generate the adjusted response text; the terms in the elderly customer friendly vocabulary library are all set with weight values, and the weight value represents the usage frequency of the term in the elderly customer group. S43. If it is determined that the adjustment direction is only sentence pattern simplification, then according to the preset simplification rule set, split the long sentences in the text to be adjusted into multiple short sentences and adjust the word order to generate the adjusted response text to make it conform to the reading habits of elderly customers, delete unnecessary modifying components, and ensure that the sentence meaning is clear. S44. If it is determined that the adjustment direction is term simplification and sentence pattern simplification, then by consulting the elderly customer friendly vocabulary library, replace the standard terms in the text to be adjusted with simplified expressions, and by applying the simplification rule set, adjust the sentence pattern and structure of the text to be adjusted after replacing the terms to generate the adjusted response text.

[0071] In this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0072] The above are only embodiments of the present invention and are not used to limit the protection scope of the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A customer VOC intelligent response method based on an e-commerce platform, characterized in that, It includes the following steps: S1. Record the language text input by the customer and the initial response text feedback by the intelligent customer service for the language text; S2. Receive the subsequent input text from the customer after receiving the initial response text; S3. According to the subsequent input text, determine whether there is an understanding obstacle. If so, generate an understanding obstacle signal and mark the initial response text as a text to be adjusted; if not, input the subsequent input text as a new language text into the intelligent customer service and return to execute step S1; S4. Adjust the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; S5. Send the adjusted response text to the corresponding customer.

2. The customer VOC intelligent response method based on the e-commerce platform according to claim 1, wherein Between step S2 and step S3, it also includes: S6. Judge the current consultation intention of the customer according to the subsequent input text, specifically including the steps: S61. Utilize a preset intention recognition model to obtain the current consultation intention of the customer by analyzing the subsequent input text; the intention recognition model is based on historical dialogue data of elderly customers and is fine-tuned and trained using the BERT model, and can recognize common consultation intentions of elderly customers and output intention confidence levels at the same time; S62. Compare the current consultation intention with the historical consultation intention to judge whether they are consistent; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence level; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence level of the current consultation intention is greater than the preset threshold and greater than the intention confidence level of the historical consultation intention, it is determined that intention drift has occurred. After inputting the subsequent input text as a new language text into the intelligent customer service, return to execute step S1, otherwise execute step S3.

3. The customer VOC intelligent response method based on the e-commerce platform according to claim 2, wherein The specific steps in step S61 include: Construct a product knowledge graph containing multiple product information; the product information includes various product attributes of various products; By parsing the subsequent input text, extract all keywords in the subsequent input text, and based on the product knowledge graph, map each keyword to a corresponding product attribute vector; Input all the product attribute vectors into the intention recognition model respectively to obtain multiple intention categories and corresponding intention confidence levels; Select the intention category with the highest confidence level as the current consultation intention of the customer.

4. The customer VOC intelligent response method based on an e-commerce platform according to claim 1, wherein The specific steps in step S3 include: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text and construct them into a text feature vector; S32. Input the text feature vector into a preset understanding obstacle detection model to obtain an understanding obstacle probability; the understanding obstacle detection model is based on historical dialogue data of elderly customers and is trained using the support vector machine (SVM) algorithm, and can recognize common interrogative, negative, and repeated expression ways of elderly customers; S33. Determine whether the probability of the comprehension barrier is greater than a preset probability threshold. If so, determine that there is an comprehension barrier, generate an comprehension barrier signal, and mark the initial response text as text to be adjusted; otherwise, determine that there is no comprehension barrier, input the subsequent input text as a new language text into the intelligent customer service, and return to execute step S1.

5. The customer VOC intelligent response method based on an e-commerce platform according to claim 4, wherein The specific steps in step S31 include: Constructing a stop word list; the stop word list includes common modal particles, auxiliary words and meaningless words; Using the stop word list to clean the subsequent input text to remove stop words in the subsequent input text, thereby obtaining a cleaned subsequent input text; Using a word segmentation algorithm to perform word segmentation processing on the subsequent input text after the cleaning to obtain multiple text segmentations; For each of the text segmentation words, determine whether it belongs to a preset question word library, a negative word library or a repeated word library; if it belongs to the question word library, mark the text segmentation word as a question word; if it belongs to the negative word library, mark the text segmentation word as a negative word; if it belongs to the repeated word library, mark the text segmentation word as a repeated word; Counting the number of question words, negative words, and repeated words in the subsequent input text, which are recorded as the number of question words, the number of negative words, and the number of repeated words, respectively; The text feature vector is constructed according to the number of question words, the number of negative words and the number of repeated words.

6. The customer VOC intelligent response method based on an e-commerce platform according to claim 1, wherein The specific steps in step S4 include: S41. Determine an adjustment method for the text to be adjusted according to the comprehension barrier signal; the adjustment method includes term simplification and sentence simplification, wherein the term simplification is used when the comprehension barrier signal indicates that the customer has difficulty understanding professional terms, and the sentence simplification is used when the comprehension barrier signal indicates that the customer has difficulty understanding long sentences; S42. If it is determined that the adjustment direction is only term simplification, then consult a preset elderly customer-friendly vocabulary library, and calculate the semantic similarity between each standard term in the text to be adjusted and the vocabulary in the elderly customer-friendly vocabulary library, and select the vocabulary with a semantic similarity higher than a preset value as the replacement vocabulary, and adjust the priority of the replacement vocabulary according to the weight value, give priority to the replacement vocabulary with the highest weight value, and generate an adjusted response text; the vocabulary in the elderly customer-friendly vocabulary library is set with a weight value, and the weight value represents the frequency of use of the vocabulary in the elderly customer group; S43. If it is determined that the adjustment direction is only sentence simplification, then according to the preset simplification rule set, the long sentence in the text to be adjusted is split into multiple short sentences and the word order is adjusted to generate an adjusted response text; S44. If the adjustment direction is determined to be term simplification and sentence simplification, the standard terms in the text to be adjusted are replaced with simplified expressions by referring to the elderly customer-friendly vocabulary library, and the sentence and structure of the text to be adjusted after the replacement of the terms are adjusted by applying the simplification rule set to generate an adjusted response text.

7. A customer VOC intelligent response system based on an e-commerce platform, characterized in that, include: A recording module, used to record the language text input by the customer and the initial response text fed back by the intelligent customer service to the customer based on the language text; A receiving module, used for receiving subsequent input text from the client after receiving the initial response text; A first judgment module, configured to judge whether there is an understanding obstacle according to the subsequent input text. If so, generate an understanding obstacle signal and mark the initial response text as a text to be adjusted; if not, input the subsequent input text into the intelligent customer service as a new language text, and return to run the recording module; A generation module, configured to adjust the text to be adjusted according to the understanding obstacle signal to generate an adjusted response text; A sending module, configured to send the adjusted response text to the corresponding customer.

8. The customer VOC intelligent response system based on an e-commerce platform according to claim 7, characterized in that, It further includes a second judgment module, which is configured to judge the current consultation intention of the customer according to the subsequent input text, and specifically execute the following steps: S61. Utilize a preset intention recognition model to obtain the current consultation intention of the customer by analyzing the subsequent input text; The intention recognition model is based on historical dialogue data of elderly customers and is fine-tuned and trained using the BERT model. It can recognize common consultation intentions of elderly customers and output intention confidence levels at the same time; S62. Compare the current consultation intention with the historical consultation intention to judge whether they are the same; the historical consultation intention is obtained by analyzing the language text through the intention recognition model, and the historical consultation intention has a corresponding intention confidence level; S63. If the current consultation intention is inconsistent with the historical consultation intention, and the intention confidence level of the current consultation intention is greater than a preset threshold and greater than the intention confidence level of the historical consultation intention, it is determined that intention drift has occurred. After inputting the subsequent input text into the intelligent customer service as a new language text, return to run the recording module, otherwise run the first judgment module.

9. The customer VOC intelligent response system based on an e-commerce platform according to claim 8, wherein When the second judgment module is used to obtain the current consultation intention of the customer by analyzing the subsequent input text using a preset intention recognition model, it executes: Construct a product knowledge graph containing multiple product information; the product information includes various product attributes of various products; By parsing the subsequent input text, extract all keywords in the subsequent input text, and based on the product knowledge graph, map each keyword to a corresponding product attribute vector; Input all the product attribute vectors into the intention recognition model respectively to obtain multiple intention categories and corresponding intention confidence levels; Select the intention category with the highest confidence level as the current consultation intention of the customer.

10. The customer VOC intelligent response system based on an e-commerce platform according to claim 7, wherein, When the first judgment module is used to judge whether there is an understanding obstacle according to the subsequent input text. If so, generate an understanding obstacle signal and mark the initial response text as a text to be adjusted; if not, input the subsequent input text into the intelligent customer service as a new language text, and return to run the recording module, it executes: S31. Extract the interrogative words, negative words, and repeated words in the subsequent input text, and construct them into a text feature vector; S32. Input the text feature vector into a pre-set comprehension obstacle detection model to obtain a comprehension obstacle probability; the comprehension obstacle detection model is based on historical elderly customer conversation data and is trained using the support vector machine (SVM) algorithm, and can identify common expressions of questions, negations, and repetitions used by elderly customers; S33. Determine whether the comprehension obstacle probability is greater than a pre-set probability threshold. If it is greater, it is determined that there is a comprehension obstacle, a comprehension obstacle signal is generated, and the initial response text is marked as a text to be adjusted. Otherwise, it is determined that there is no comprehension obstacle, the subsequent input text is used as a new language text and input into the intelligent customer service, and the operation of the recording module is returned.

Images