Dynamic context window customer service quality inspection method and system based on large model

Abstract

The invention discloses a dynamic context window customer service quality inspection method based on a large model, and the method comprises the following steps: S1, generating a dynamic window: scoring the dialogue round of a customer service and a user, dynamically updating and maintaining a context window according to a scoring result, and carrying out the context reconstruction of a dialogue in the window; s2, multi-dimensional quality inspection: inputting the dialogue text in the dynamic window into a large model, and outputting a preset field through a customized prompt trigger model; s3, clustering attribution: carrying out semantic clustering on related contents output by the large model by adopting a kmeans algorithm and BERT vectorization, and then generating a general description and an operable suggestion for each clustering result through prompt; and S4, result output and application: generating a structured json result containing a multi-dimensional quality inspection result and a clustering result, wherein the structured json result is used for api calling or visual platform display. The method has the advantages that efficient, accurate and multi-dimensional customer service quality inspection can be achieved, and the service quality can be improved and overcome.

Description

Technical Field

[0001] The present invention belongs to the technical field of customer service quality inspection, and in particular relates to a dynamic context window customer service quality inspection method and system based on a large model. Background Art

[0002] In today's digital age, customer service serves as a crucial bridge between businesses and their customers, and its quality directly impacts customer satisfaction and corporate image. Customer service quality inspection, a key step in the customer service process, is crucial for improving a company's service levels by analyzing historical conversation records to evaluate service quality, identify customer needs, and optimize service processes. Traditional customer service quality inspection technologies primarily rely on rule engines or statistical-based natural language processing (NLP) methods, but these methods have numerous limitations and are unable to meet the increasingly complex needs of customer service quality inspection: For example, using a static context window is a traditional method. Existing methods usually use a fixed-length context window, for example, only considering the last 10 rounds of conversation. This fixed window approach cannot dynamically adapt to long conversation scenarios or conversations with complex logic. In long conversations, key information may be distributed at different stages of the conversation. The fixed window can easily miss important content, such as the gradual changes in user emotions and the trajectory of problem solving, which can lead to misjudgment of customer service quality. For example, in an after-sales complaint conversation lasting 20 rounds, the user expressed strong dissatisfaction in the fifth round. Since the fixed window only retains the last 10 rounds of conversation, this key emotional information will be excluded, making it impossible for quality inspectors to fully understand the customer's emotional changes, thereby making an inaccurate assessment of the customer service's emotional management capabilities.

[0003] Single-dimensional analysis is another major shortcoming of traditional quality inspection methods. Traditional quality inspections often only focus on basic indicators such as service attitude and problem-solving results, and lack in-depth analysis of multiple dimensions. For example, when users’ emotions change from anger to satisfaction, traditional methods cannot deeply explore the reasons behind it. It is difficult to analyze whether the customer service’s communication skills played a key role or whether the smooth resolution of the problem changed the user’s attitude. At the same time, the assessment of customer service capabilities is also relatively general and cannot be broken down into specific aspects such as communication skills, professional knowledge, and proactive service awareness. This makes it difficult for companies to effectively train and improve weak links in customer service. In addition, traditional methods are unable to tap into potential risks of products, such as the clustering of causes of customer churn, which is not conducive to companies discovering product problems in a timely manner and making improvements.

[0004] The lack of clustering and pattern mining is also a major shortcoming of traditional technology. Existing technology does not deeply aggregate quality inspection results and cannot extract common problems from massive conversation data. For example, in a large number of pre-sales consultation conversations, there may be multiple customers asking about a functional defect of the same product, but traditional methods cannot cluster and analyze these scattered consultation information, and thus cannot discover this high-frequency consultation hotspot. Similarly, for repetitive service defects, such as the frequent unclear process explanations when customer service handles refund issues, traditional methods are also difficult to identify. This leads to the lack of systematic and targeted optimization suggestions proposed by the company, and is unable to fundamentally improve the quality of customer service.

[0005] To sum up, traditional customer service quality inspection technology can no longer meet the modern enterprises' pursuit of high-quality customer service. A new and more efficient customer service quality inspection method and system is urgently needed to solve the above problems. Summary of the Invention

[0006] In order to address the deficiencies in the prior art, the present invention provides a large-model-based dynamic context window customer service quality inspection method and system. The large-model-based dynamic context window customer service quality inspection method and system can efficiently, accurately, and multi-dimensionally inspect customer service quality, thereby improving and overcoming service quality.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions: The dynamic context window customer service quality inspection method based on a large model includes the following steps: S1. Dynamic Window Generation: Based on the original conversation flow between the customer service representative and the user, the conversation turns between the customer service representative and the user are scored. The context window is dynamically updated and maintained based on the scoring results, and the conversation within the window is reconstructed. S2. Multi-dimensional quality inspection: Input the conversation text in the dynamic window into the large model, trigger the model to output preset fields through customized prompts, and implement analysis of basic indicators, customer dimensions, customer service dimensions, and product dimensions; S3. Clustering Attribution: Use the kmeans algorithm and BERT vectorization to semantically cluster the relevant content output by the large model, and then generate a summary description and actionable suggestions for each clustering result through prompts; S4. Result output and application: Generate structured JSON results containing multi-dimensional quality inspection results and clustering results for API calls or visualization platform display.

[0008] Preferably, the dialogue round scoring in S1 includes a question density score, an emotion fluctuation score, a structural key node score, and an interaction intensity score, and the dialogue round scoring is performed by calculating a critical score through the question density score, the emotion fluctuation score, the structural key node score, and the interaction intensity score. The question density score is used to determine whether the dialogue contains clear user demands, feedback, or consultation keywords, the emotion fluctuation score is used to determine whether the customer's emotions have changed significantly, the structural key node score is used to determine whether the dialogue is in a critical stage, and the interaction intensity score is used to determine whether the current round has triggered intensive interactions. The critical score calculation formula is: Critical score = 1× Question Density Score + 2× Mood Swing Score+ 3×Structural Key Node Score+ 4×interaction intensity score, where 1~ 4 is the tuning weight.

[0009] Preferably, the capacity of the context window queue in S1 is determined by the maximum token length limit, and the dynamic update window adopts a retention strategy, a culling strategy, and a cold start strategy; Retention strategy: prioritize retaining the highest-scoring rounds of dialogue; Elimination strategy: If the total number of tokens in a window exceeds the limit, the round with the lowest cumulative score will be eliminated; Cold start strategy: For newly entered conversations, the first N rounds are retained by default.

[0010] Preferably, the basic indicators in S2 include dialogue scenarios, solution status and dialogue core content, wherein the dialogue scenario is used to distinguish pre-sales or after-sales scenarios, and the solution status is used to determine whether the demands raised by the customer in the dialogue are resolved.

[0011] Preferably, the customer dimensions in S2 include customer emotion changes, feedback issues and consultation hotspots, wherein customer emotion changes include the start or end score and reason.

[0012] Preferably, the customer service dimensions in S2 include quantitative customer service capabilities, reasons for service satisfaction and improvement suggestions, wherein the quantitative customer service capabilities include scores and reasons for five dimensions: attitude, communication, professionalism, proactivity and emotional management.

[0013] Preferably, the product dimension in S2 includes reasons for abandoning purchase and specific improvement suggestions.

[0014] Preferably, the S3 performs semantic clustering on the relevant content output by the large model, including customer consultation hotspots, customer feedback issues, reasons for satisfaction, reasons for abandoning purchases, reasons for changes in customer emotions, and customer service capabilities.

[0015] Preferably, the summary description and actionable suggestions generated in S3 include specific scenarios.

[0016] The present invention also discloses a large-model-based dynamic context window customer service quality inspection system, comprising a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the above-mentioned large-model-based dynamic context window customer service quality inspection method is implemented.

[0017] By adopting the above technical solution, the present invention has the following beneficial effects: (1) Dynamic context has obvious advantages. This application can ensure that key information in the conversation is not missed through the design of a dynamic context window, thereby making the quality inspection indicators more accurate. Compared with the traditional fixed context window, the dynamic window can be flexibly adjusted according to the importance of the conversation content, retaining those information rounds that are crucial for evaluating customer service quality and understanding customer needs. At the same time, the dynamic window can reduce excessive input of irrelevant context, reduce the reasoning delay of large models, save computing resources, and improve quality inspection efficiency. For example, in a long conversation, the dynamic window can automatically filter out key information rounds, avoiding the waste of resources and inefficiency caused by large models processing a large amount of redundant information, and can accurately and efficiently perform customer service quality inspection; (2) The multi-dimensional fine-grained analysis capability is outstanding. The present invention realizes the independent scoring and cause analysis of the five dimensions of customer service capabilities (attitude, communication, professionalism, proactivity and emotional management), providing a precise target for customer service training. The traditional method is relatively general in its assessment of customer service capabilities and cannot specifically point out the deficiencies in customer service. The present invention can deeply analyze the performance of each dimension, such as accurately identifying that "lack of professional ability" is specifically manifested as "inability to explain product technical parameters", so that enterprises can carry out targeted training for these specific problems and quickly improve the professional quality of customer service. In addition, multi-dimensional analysis can also comprehensively explore user needs and product risks, provide a more comprehensive basis for the company's product improvement and service optimization, and conduct customer service quality inspection in multiple dimensions; (3) Intelligent clustering and decision support play a significant role. The automatic clustering and summarization function of this application significantly reduces the cost of manual analysis. By semantically clustering massive quality inspection results, common problems and patterns can be quickly discovered. For example, through cluster analysis, it is found that multiple customers have questions about a specific function of a product. The company can promptly organize customer service to conduct relevant training or optimize product descriptions. At the same time, the generated actionable suggestions are highly practical and targeted, providing strong decision support for the company's operations and training, helping the company to continuously improve customer service quality and customer satisfaction. In summary, the present invention has the advantages of efficient, accurate and multi-dimensional customer service quality inspection, which can improve and overcome service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow diagram of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0020] The components of the embodiments of the present invention generally described and shown in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention.

[0021] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0022] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0024] Example 1 In this embodiment, a dynamic context window customer service quality inspection method and system based on a large model is proposed, which is applicable to e-commerce customer service, intelligent customer service systems, corporate training and other fields. In the field of e-commerce customer service, it can conduct real-time quality inspection of pre-sales consultations (such as answers to promotional activities) and after-sales complaints (such as return process processing), identify customer service language loopholes and product after-sales risks. In the intelligent customer service system, it can be integrated into the customer service robot to realize automatic quality inspection and strategy optimization (such as dynamically adjusting the robot's reply language). In the field of corporate training, it can generate customized training courses based on the results of ability clustering (such as "lack of communication skills" is concentrated in "multi-round questioning scenarios").

[0025] The present invention aims to solve many problems existing in existing customer service quality inspection technology, specifically as follows: (1) Dynamic context modeling is the primary technical problem that the present invention needs to solve. When a large model performs quality inspection on a customer service conversation, the input length of the large model is usually limited by the size of its context window. If the input conversation content is lengthy and complex, or contains a large amount of information irrelevant to the quality inspection, it will not only waste valuable computing resources and increase processing costs, but may also distract the large model and prevent it from focusing on key information, thereby affecting the accuracy of the quality inspection results. Therefore, designing a method that can dynamically retain the most critical information rounds in the conversation and constructing an efficient, low-redundancy, and high-information-density context window for subsequent quality inspection calls of the large model is an important goal of the present invention. (2) Multi-dimensional intelligent quality inspection is another problem faced by the present invention. How to use a large model to achieve a multi-dimensional and fine-grained evaluation of customer service is an urgent problem to be solved. Specifically, it is necessary to use a large model to conduct a detailed evaluation of customer service capabilities such as service attitude, professional ability, and emotional management. At the same time, it is also necessary to be able to deeply explore the potential needs of users and possible risks of products. Traditional single-dimensional analysis can no longer meet the needs of enterprises for a comprehensive understanding of customer service quality. Therefore, realizing multi-dimensional intelligent quality inspection and providing enterprises with more comprehensive and in-depth quality inspection information is another important task of the present invention. (3) Semantic-level cluster analysis is also a technical problem that the present invention needs to solve. How to effectively semantically cluster the quality inspection results based on the generation capability of large models and output actionable improvement suggestions is the key to improving the practicality of customer service quality inspection. By performing semantic-level cluster analysis on massive quality inspection results, common problems and patterns can be discovered, and targeted improvement measures can be proposed to help companies optimize service processes and customer service training systems. However, to achieve semantic-level cluster analysis, it is necessary to solve problems such as the accurate calculation of text semantic similarity and the effective interpretation of clustering results.

[0026] like Figure 1As shown, in one embodiment of the present invention, the dynamic context window customer service quality inspection method based on a large model of the present invention includes the following steps: S1. Dynamic Window Generation: Based on the original conversation flow between the customer service representative and the user, the conversation turns between the customer service representative and the user are scored. The context window is dynamically updated and maintained based on the scoring results, and the conversation within the window is reconstructed. It is understandable that the core of the dynamic window generation step in S1 is the three steps of dialogue turn scoring, window update and maintenance, and context reconstruction; In the dialogue round scoring phase, each dialogue round (including the dialogue between customer service and users) is scored, and a "criticality score" is calculated based on multiple dimensions to quantify the importance of this dialogue round in the entire dialogue process. Specifically, the dialogue round scoring in S1 includes the question density score, the emotional fluctuation score, the structural key node score, and the interaction intensity score. The dialogue round scoring is calculated based on the question density score, the emotional fluctuation score, the structural key node score, and the interaction intensity score. The question density score is one of the important scoring dimensions. It mainly determines whether the conversation contains clear user demands, feedback, or consultation keywords, such as "cannot open", "how to refund", "product parameters", etc. If a conversation contains multiple such keywords, or the keywords included are crucial for understanding user needs and evaluating customer service, the question density score of this conversation will be relatively high. For example, if a user clearly asks "When will the product I purchased be shipped?" in a certain round, this conversation contains key consultation keywords and its question density score will be high. For the question density score, we can pre-build a vocabulary containing common user demands, feedback, and consultation keywords, such as "refund", "shipping", "quality", "price", etc. When analyzing a certain round of conversation, we count the number and importance of the keywords in the vocabulary that appear in this round of conversation. For example, if the keywords "refund" and "quality issue" appear in a certain round of conversation, and "quality issue" has a higher weight in the vocabulary, then the question density score of this round will be correspondingly high. The question density score can be set to 0-10 points and calculated based on the number and weight of keywords. The emotion fluctuation score focuses on whether the customer's emotions have changed significantly, such as from calm to angry, from dissatisfied to satisfied, from positive to indifferent, etc. By analyzing the customer's language expression, modal particles, punctuation marks, etc., the customer's emotion fluctuations are judged. If the customer's emotions have undergone a drastic positive or negative change in a certain round of conversation, the emotion fluctuation score of that round of conversation will increase accordingly. For example, if the customer has been expressing dissatisfaction before, and suddenly says "OK, I understand, thank you" after a certain round of conversation, this indicates that the customer's emotions have changed from dissatisfaction to satisfaction, and the emotion fluctuation score of that round will be higher. The calculation of the emotion fluctuation score can be Using a sentiment analysis model, we first analyze the sentiment tendency of the customer's text in each round of conversation to obtain positive, negative, or neutral sentiment labels and assign corresponding sentiment intensity values ​​(e.g., 0-10, with 10 being the most positive). We then calculate the difference in sentiment intensity between the current round and the previous round, and determine the sentiment fluctuation score based on the difference. A large absolute value of the difference indicates a significant change in the customer's mood, resulting in a higher sentiment fluctuation score (e.g., 0-10). For example, if the customer's sentiment intensity in the previous round was 3 (negative) and the current round was 8 (positive), the difference is 5 points, and the sentiment fluctuation score can be determined as 7. The Structural Key Node score is used to determine whether a conversation is at a critical stage, such as the opening phase, the initial question expression phase, the problem-solving phase, and the closing phase. These critical phases often contain important information and are crucial for understanding the overall context of the conversation and evaluating customer service quality. For example, if a user clearly states their core concern during the initial question expression phase, the Structural Key Node score for that round of conversation will be higher. The judgment of the Structural Key Node score needs to be considered in conjunction with the conversation's development. The opening phase can be determined by the conversation's starting marker (e.g., "Hello, how may I help you?"). The initial question expression phase occurs when the user clearly states their core concern for the first time. The problem-solving phase occurs when the customer service representative provides a solution or the user indicates the problem has been resolved. The closing phase occurs when the conversation ends (e.g., "Thank you for your inquiry, goodbye"). For conversations at these critical stages, the Structural Key Node score can be set to a higher value (e.g., 0-10 points, with 8-10 points for critical stages and 1-5 points for non-critical stages). The interaction intensity score mainly examines whether the current round has triggered intensive interactions, such as repeated confirmations and quick responses from both parties. If, after a round of dialogue, the two parties exchange and confirm information multiple times, it indicates that the dialogue has triggered strong interactions, and its interaction intensity score will be higher. For example, if customer service raises a key piece of information that requires user confirmation in a round, the user then responds, and customer service asks further questions, the interaction intensity score of such a round will be relatively high. The interaction intensity score is mainly determined by the frequency of interactions after the round of dialogue. If, after a round of dialogue, the two parties have multiple exchanges of information and responses within a relatively short period of time, it indicates that the dialogue has triggered intensive interactions, and the interaction intensity score is high (e.g., 0-10 points). For example, after a round of dialogue, the two parties have three or more rounds of dialogue within 5 minutes, the interaction intensity score of this round can be set as 8 points. Specifically, the criticality score calculation formula is: Critical score = 1× Question Density Score + 2× Mood Swing Score+ 3×Structural Key Node Score+ 4×interaction intensity score, where 1~ 4 is the weight for tuning, and it is fine-tuned according to the business scenario. For example, in the after-sales complaint scenario, the customer's emotional fluctuations may be more important, so the weight can be appropriately increased. 2 weight; in the pre-sales consultation scenario, the density of questions may be the key, which can improve With a weight of 1, assuming 1=0.3, 2=0.3, 3=0.2, 4=0.2, the question density score of a round of dialogue is 8 points, the emotional fluctuation score is 6 points, the structural key node score is 7 points, and the interaction intensity score is 5 points. Then the criticality score of this round = 0.3×8+0.3×6+0.2×7+0.2×5=2.4+1.8+1.4+1=6.6 points.

[0027] During the window update and maintenance phase, a sliding "context window queue" is maintained. Its capacity is determined by the maximum token length limit (e.g., Mtokens). This can be set based on the processing capability of the large model used, for example, to 2048 tokens. Furthermore, multiple strategies are used to dynamically update the window during the window update and maintenance phase to ensure that the information within the window is both critical and within the processing capability of the large model. In other words, the capacity of the context window queue in S1 is determined by the maximum token length limit, and the dynamic window update adopts a retention strategy, a culling strategy, and a cold start strategy. The retention strategy prioritizes the highest-scoring rounds. By sorting the criticality scores of each round, the rounds with higher scores are included in the context window to ensure that the window contains the most important information in the conversation. Specifically, the criticality scores of all rounds can be sorted from high to low, and the rounds with higher scores are prioritized for inclusion in the window. For example, if there are 15 rounds of conversation and the window can accommodate a maximum of 10 rounds, the top 10 rounds are selected. The elimination strategy eliminates the round with the lowest cumulative score when the total number of tokens in a window exceeds the maximum limit. This strategy ensures that the input content of the large model does not exceed its processing limit while ensuring the criticality of the information, thus avoiding resource waste and distraction. Specifically, when the total number of tokens in a window exceeds the maximum limit, the elimination strategy is triggered. The cumulative score of each round of dialogue in the current window (i.e., the criticality score of that round) is calculated, and the round with the lowest score is eliminated until the total number of tokens in the window is within the limit. The cold start strategy primarily targets newly entered conversations. The first N rounds are retained by default to avoid missing critical information from the opening phase. For example, for a new customer service conversation, the first five rounds are retained in the window by default, regardless of their scores. This is because information from the opening phase is crucial for understanding the context and topic of the entire conversation. Specifically, the value of N can be set based on the actual business scenario. For example, if N=5, the first five rounds of a new conversation are retained in the window by default, regardless of their criticality scores.

[0028] The context reconstruction phase concatenates the selected turns into the final context input in their original order, while preserving the format information (speaker, timestamp, original content), and constructs a prompt format suitable for large models. For example: PlainText: [2025-01-01 21:00:00] [User] Hello, my order has been placed for two days and it has not been shipped yet? [2025-01-01 21:01:00] [Customer Service] Hello, please wait, I will check for you; [2025-01-01 21:02:00] [User] I have encountered this situation before. Will I have to delay again? [2025-01-01 21:03:00] [Customer Service] The system currently shows that the order is expected to be shipped tomorrow. We are very sorry. The above format can clearly present the context of the conversation and related information, making it easier for large models to analyze.

[0029] S2. Multi-dimensional quality inspection: Input the conversation text in the dynamic window into the large model, trigger the model to output preset fields through customized prompts, and implement analysis of basic indicators, customer dimensions, customer service dimensions, and product dimensions; The multi-dimensional quality inspection step is to achieve accurate quality inspection. The conversation text in the dynamic window of S1 is input into the large model. The model is triggered by customized prompts to output preset fields, realizing comprehensive analysis of multiple dimensions. The design of customized prompts is crucial. The large model must be clearly informed of the dimensions to be analyzed and the output format. For example, "Please conduct a quality control analysis on the following customer service conversations and output the following fields: conversation scenario (pre-sales / post-sales), resolution status (yes / no), core content of the conversation, changes in customer sentiment (start rating and reason, end rating and reason), feedback issues, hot topics, customer service capabilities (attitude, communication, professionalism, proactivity, and emotion management ratings and reasons), reasons for service satisfaction and suggestions for improvement, reasons for abandoning a purchase and specific suggestions for improvement." The basic indicators in S2 include conversation scenarios, resolution status, and core content of the conversation. The conversation scenarios are used to distinguish between pre-sales and post-sales scenarios. Pre-sales scenarios may involve product introductions, answers to promotional questions, etc., while post-sales scenarios may include return process processing, product quality complaints, etc. The resolution status is used to determine whether the customer's request in the conversation has been resolved. For example, if a customer inquires about a product refund, if the customer service successfully processes the refund for the customer, the resolution status is "yes". The core content of the conversation is to extract the core theme and key information of the entire conversation, such as "the customer inquires about product logistics information and estimated delivery time"; The customer dimension in S2 includes customer sentiment changes, feedback issues, and hot topics. Customer sentiment changes include the start or end score and the reason. Specifically, customer sentiment changes should not only provide the sentiment scores at the beginning and end (e.g., 1-10, with 1 being the most negative and 10 being the most positive), but also explain the reasons for the scores. For example, the start sentiment score is 3 because the customer was dissatisfied because the product was not delivered on time, and the end sentiment score is 8 because the customer service promptly resolved the issue and apologized. Feedback issues refer to the dissatisfaction and opinions about the product or service raised by the customer during the conversation, such as "damaged product packaging." Hot topics are the topics that customers frequently inquire about, such as "the expiration date of a certain product's promotion." The customer service dimension in S2 includes quantitative customer service capabilities, reasons for service satisfaction, and improvement suggestions. Quantitative customer service capabilities include scores and reasons for the five dimensions of attitude, communication, professionalism, proactivity, and emotional management. For example, an attitude score of 4 points is due to the customer service using polite language throughout the process. A professionalism score of 3 points is due to the customer service's unclear explanation of some product parameters. Reasons for service satisfaction are analyzed based on the specific factors that led to customer satisfaction or dissatisfaction with the service. Improvement suggestions are specific improvement measures proposed for deficiencies in customer service, such as "It is recommended that customer service strengthen their learning of product parameters." The product dimension in S2 includes the reasons for abandoning purchase and specific improvement suggestions. Specifically, the reasons for abandoning purchase may include that the product price is too high, the functions do not meet the needs, the logistics speed is slow, etc. The specific improvement suggestions are based on these reasons, such as "to address the price issue, more preferential packages can be launched; to address the logistics issue, the logistics distribution plan can be optimized."

[0030] S3. Clustering Attribution: Use the kmeans algorithm and BERT vectorization to semantically cluster the relevant content output by the large model, and then generate a summary description and actionable suggestions for each clustering result through prompts; The clustering and attribution step in S3 is used to mine patterns from massive quality inspection results and provide targeted improvement directions for enterprises. It includes two steps: semantic clustering using a clustering algorithm and secondary generation of a large model. S3 semantically clusters the relevant content output by the large model, including customer consultation hotspots, customer feedback issues, reasons for satisfaction, reasons for abandoning purchases, reasons for changes in customer sentiment, and customer service capabilities. The summary descriptions and actionable suggestions generated in S3 include specific scenarios. Specifically, the semantic clustering phase of the clustering algorithm uses the kmeans algorithm and BERT vectorization to perform semantic clustering on the output of the large model, including customer consultation hotspots, customer feedback issues, reasons for satisfaction, reasons for abandoning purchases, reasons for customer sentiment changes, and customer service capabilities (including attitude, communication, professionalism, proactivity, and emotional management). This is combined with the results of historical conversations. Unlike traditional clustering based on word frequency, this method considers the semantic similarity of text and can more accurately classify issues. BERT vectorization converts text into a vector representation with semantic information, so that texts with similar semantics are closer in the vector space. For example, the two texts "slow arrival of goods" and "too slow logistics" have slightly different wording but similar semantics. After BERT vectorization, their vector representations will be very close. The kmeans algorithm clusters these vectors based on the distance between them, grouping semantically similar content into one category. Specifically, the text content in the quality inspection results output by the large model (such as customer consultation hotspots, feedback issues, etc.) is batch-fed into the BERT model to obtain the vector representation of each text. The BERT model can use a pre-trained Chinese BERT model, such as bert-base-chinese; The kmeans algorithm clustering is to set the number of clusters k of the K-means algorithm based on the actual number of quality inspection results and business needs. For example, for customer consultation hotspots, k can be set to 5. The vector representation of the text is input into the K-means algorithm for clustering to obtain different clusters; During the secondary generation phase of the large model, a summary description (such as "summarized_consult") and actionable suggestions (such as "summarized_suggestion") are generated for each clustering result through prompts to ensure that the suggestions include specific scenarios; For example, for the cluster result "Logistics delay," the summary description might be "Multiple customers reported that product logistics delivery was delayed, affecting the customer experience," and the actionable suggestion might be "When users inquire about logistics delays, customer service should proactively provide compensation plans, such as coupons or small gifts, and explain the cause of the delay and the estimated delivery time to the user." More specifically, for each cluster, a corresponding prompt is constructed, such as "Please summarize the common content of the following customer consultation hotspots and provide specific improvement suggestions: [hot consultation content in the cluster]." This prompt is input into the large model to generate a summary description and actionable suggestions. For example, for a cluster containing content such as "logistics is too slow" and "delivery time is long", the summary description generated by the large model may be "Customers generally report that product logistics delivery is slow and delivery time is long." The actionable suggestion may be "Optimize logistics partners to improve delivery efficiency, inform customers of the approximate arrival time in advance, and proactively explain the reasons and provide compensation for orders that may be delayed." Through the above steps, we can achieve analysis in multiple dimensions, including cluster summary of consultation hot spots, cluster summary of user feedback, cluster summary and improvement suggestions of satisfaction reasons, cluster summary and improvement suggestions of reasons for abandoning purchases, cluster summary of reasons for customer emotion changes, and cluster summary of customer service capabilities.

[0031] S4. Result output and application: Generate structured JSON results containing multi-dimensional quality inspection results and clustering results for API calls or visualization platform display; Specifically, a structured JSON output is generated, containing multi-dimensional quality inspection results and clustering results (clustering summaries and improvement suggestions, as well as cluster IDs based on historical data). This structured JSON output can be easily used for API calls, integrated with other enterprise systems, and displayed on a visualization platform, enabling enterprise managers to intuitively understand customer service quality inspection status.

[0032] The present invention also discloses a large-model-based dynamic context window customer service quality inspection system, comprising a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the above-mentioned large-model-based dynamic context window customer service quality inspection method is implemented.

[0033] This embodiment does not impose any formal restrictions on the shape, material, structure, etc. of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are within the scope of protection of the technical solution of the present invention.

Claims

1. A dynamic context window customer service quality inspection method based on a large model, characterized by: The following steps are involved: S1. Dynamic Window Generation: Based on the original conversation flow between the customer service representative and the user, the conversation turns between the customer service representative and the user are scored. The context window is dynamically updated and maintained based on the scoring results, and the conversation within the window is reconstructed. S2. Multi-dimensional quality inspection: Input the conversation text in the dynamic window into the large model, trigger the model to output preset fields through customized prompts, and implement analysis of basic indicators, customer dimensions, customer service dimensions, and product dimensions; S3. Clustering Attribution: Use the kmeans algorithm and BERT vectorization to semantically cluster the relevant content output by the large model, and then generate a summary description and actionable suggestions for each clustering result through prompts; S4. Result output and application: Generate structured JSON results containing multi-dimensional quality inspection results and clustering results for API calls or visualization platform display.

2. The large model-based dynamic context window customer service quality inspection method according to claim 1, characterized in that: The dialogue round scoring in S1 includes a question density score, an emotion fluctuation score, a structural key node score, and an interaction intensity score. The dialogue round scoring is calculated by using the question density score, the emotion fluctuation score, the structural key node score, and the interaction intensity score. The question density score is used to determine whether the dialogue contains clear user demands, feedback, or consultation keywords. The emotion fluctuation score is used to determine whether the customer's emotions have changed significantly. The structural key node score is used to determine whether the dialogue is in a critical stage. The interaction intensity score is used to determine whether the current round has triggered intensive interactions. The critical score calculation formula is: Critical score = 1× Question Density Score + 2× Mood Swing Score+ 3×Structural Key Node Score+ 4×interaction intensity score, where 1~ 4 is the tuning weight.

3. The large model-based dynamic context window customer service quality inspection method according to claim 1 is characterized by: The capacity of the context window queue in S1 is determined by the maximum token length limit, and the dynamic update window adopts the retention strategy, the elimination strategy and the cold start strategy; Retention strategy: prioritize retaining the highest-scoring rounds of dialogue; Elimination strategy: If the total number of tokens in a window exceeds the limit, the round with the lowest cumulative score will be eliminated; Cold start strategy: For newly entered conversations, the first N rounds are retained by default.

4. The large model-based dynamic context window customer service quality inspection method according to claim 1 is characterized by: The basic indicators in S2 include dialogue scenarios, resolution status, and core dialogue content. The dialogue scenarios are used to distinguish pre-sales and post-sales scenarios, and the resolution status is used to determine whether the demands raised by customers in the dialogue have been resolved.

5. The large model-based dynamic context window customer service quality inspection method according to claim 1 is characterized by: The customer dimension in S2 includes customer sentiment changes, feedback issues and consultation hotspots, wherein customer sentiment changes include the start or end score and reasons.

6. The large model-based dynamic context window customer service quality inspection method according to claim 1, characterized in that: The customer service dimensions in S2 include quantitative customer service capabilities, reasons for service satisfaction and improvement suggestions, among which quantitative customer service capabilities include scores and reasons for five dimensions: attitude, communication, professionalism, proactivity and emotional management.

7. The large model-based dynamic context window customer service quality inspection method according to claim 1, characterized in that: The product dimension in S2 includes the reasons for abandoning purchase and specific improvement suggestions.

8. The large model-based dynamic context window customer service quality inspection method according to claim 1, characterized in that: The S3 performs semantic clustering on the relevant content output by the large model, including customer consultation hotspots, customer feedback issues, reasons for satisfaction, reasons for abandoning purchases, reasons for changes in customer emotions, and customer service capabilities.

9. The large model-based dynamic context window customer service quality inspection method according to claim 1, characterized in that: The summary description and actionable suggestions generated in S3 contain specific scenarios.

10. A dynamic context window customer service quality inspection system based on a large model, characterized by: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the dynamic context window customer service quality inspection method based on a large model as described in any one of claims 1 to 9 is implemented.

Images