An ai customer service digital employee system and method applied to the detection industry

Abstract

The application discloses an AI customer service digital employee system and method applied to the detection industry, specifically to the intelligent field of detection industry customer service, and comprises two parts of system architecture and implementation method.The system comprises data acquisition, preprocessing, fine processing, mathematical model analysis, hierarchical output, knowledge base and data feedback modules, which are respectively used for multi-channel original data acquisition marking, data standardization processing, structured feature data set generation, double model quantitative analysis, three-level progressive response output, professional data storage and closed-loop optimization.The method steps are as follows: S1, collecting multiple types of data and classifying and marking; S2, processing data to obtain a standardized data set; S3, generating a structured feature data set; S4, model analysis and confidence test; S5, hierarchical output according to the confidence; S6, knowledge base updating and model optimization.The application solves the problems of poor traditional customer service data integration, low matching precision and rigid response, and improves the efficiency and professionalism of the detection industry customer service.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent customer service in the detection industry, and more particularly, to an AI customer service digital employee system and method applied to the detection industry. BACKGROUND

[0002] With the rapid development of the detection industry, users' demand for consultation of detection services is increasing, and the consultation content covers multiple dimensions such as detection project details, detection process specifications, detection cost standards, detection report interpretation, and detection qualification inquiry. The traditional manual customer service mode has problems such as low response efficiency, high labor cost, serious queuing during peak consultation period, and insufficient standardization of replies, which cannot meet the users' demand for fast, accurate, and professional consultation services.

[0003] The current industry customer service mainly relies on the combination of the traditional manual customer service mode and the basic AI response system: the manual customer service is responsible for handling complex consultation and user interface, and the basic AI system responds to standardized problems through simple keyword matching.

[0004] However, in actual use, there are still some shortcomings, such as the existing technology cannot integrate multi-channel heterogeneous data, the data lacks unified classification marking, and the detection industry professional terms are not standardized, non-standard expressions can lead to information bias; the voice data is simply transcribed without combining emotion feature analysis, which cannot adapt to the user's emotional state to adjust the response mode, and the basic AI system relies on single keyword matching logic, without considering multi-dimensional factors such as user industry, historical consultation frequency, and interaction scene, which makes it difficult to accurately identify the user's core demand, and often leads to mismatch between reply and demand, and the existing technology lacks a hierarchical response mechanism, which requires a complex matching process to clarify the demand, and the response efficiency is low; fuzzy demand lacks effective guidance means and cannot accurately complete the information; when high-complexity demand is transferred to manual, the key data such as user basic information and consultation context are not completely transmitted, which leads to repeated communication of manual customer service, poor service experience, and finally the knowledge base of the existing technology relies on manual input, the response lags behind the changes of industry standards, detection projects, etc., the AI model parameters are fixed, and cannot be dynamically adjusted combined with user feedback and changes in consultation data, which reduces the adaptability after long-term use. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an AI customer service digital employee system and method applied to the detection industry, which solves the problems in the background art by the following scheme.

[0006] To achieve the above object, the application provides the following technical scheme: an AI customer service digital employee method applied to the detection industry, comprising the following steps: S1, data acquisition: collecting original data of customer service scenes in the detection industry through a multi-channel interface, and classifying and marking various collected original data, including marking data types, collection times, collection channels, and user identities, to form marked original data;

[0007] S2, data preprocessing: processing the marked original data through text data processing, voice data processing, and structured data processing to obtain a standardized preprocessing data set;

[0008] S3, data reprocessing: processing the standardized preprocessing data set through semantic word segmentation and keyword extraction, keyword classification and marking, and data correlation mapping to output a structured feature data set;

[0009] S4, mathematical model data analysis: performing multivariate linear regression model construction and calculation, Bayesian probability model calculation, and confidence test based on the structured feature data set to obtain optimal matching results and confidence;

[0010] S5, hierarchical output: outputting first-level output, second-level output, and third-level output based on confidence and optimal matching;

[0011] S6, closed-loop optimization: optimizing the output results, including updating a knowledge base and optimizing a model.

[0012] An AI customer service digital employee method applied to the detection industry, comprising a data acquisition module, a data preprocessing module, a data reprocessing module, a mathematical model analysis module, a hierarchical output module, a knowledge base module, and a data feedback module;

[0013] The data acquisition module is configured to collect original data of customer service scenes in the detection industry through a preset multi-channel interface, wherein the original data comprises detection business consultation text data, user voice interaction recording data, historical customer service dialogue data, detection industry knowledge base structured data, and user basic information data; meanwhile, various collected original data is classified and marked, the marking content comprises data types, collection times, collection channels, and user identities, and the marked original data is transmitted to the data preprocessing module;

[0014] The data preprocessing module is configured to receive the marked original data transmitted by the data acquisition module, and perform standardized processing according to data types, specifically including text data cleaning and term standardization, voice data transcription and emotion marking, structured data verification and format unification, to form a standardized preprocessing data set, and transmit the standardized preprocessing data set to the data reprocessing module;

[0015] The data fine processing module is used for receiving a standardized pre-processing data set, performing semantic word segmentation and keyword extraction, keyword classification labeling, data correlation mapping, establishing a correlation data of "user basic information-advisory text-keyword category" indexed by a user identifier, outputting a structured feature data set, and transmitting the structured feature data set to the mathematical model analysis module;

[0016] The mathematical model analysis module is used for receiving the structured feature data set, performing quantitative analysis by adopting a combination of a multiple linear regression model and a Bayesian probability model, calculating an advisory intention matching degree and a posterior probability of a reply content after correction, determining a demand matching result type through confidence testing, and transmitting the matching result and the corresponding confidence to the hierarchical output module;

[0017] The hierarchical output module is used for receiving the matching result and the confidence transmitted by the mathematical model analysis module, outputting a response in a three-level progressive mode, and transmitting related data of the output response to the data feedback module;

[0018] The knowledge base module is used for storing detection industry professional data and standardized reply contents, including detection project names and technical parameters, detection process steps and operation specifications, detection fee pricing standards, detection report index meanings and interpretation rules, detection agency qualification certification information, detection standard regulations, and manually reviewed customer service reply contents, to provide data support for the hierarchical output module;

[0019] The data feedback module is used for receiving the output response related data transmitted by the hierarchical output module, reviewing the manually reviewed customer service reply contents, storing the reviewed contents in the knowledge base module, updating the "keyword-reply content" corresponding relationship data, and simultaneously associating the user feedback results with the related data as new training samples and transmitting the new training samples to the data fine processing module for model parameter adjustment, to realize system closed-loop optimization.

[0020] Technical effects and advantages of the present application:

[0021] 1. The present application supports comprehensive collection of multi-channel heterogeneous data, uniformly marks data types, collection times and other key dimensions, realizes terminology standardization, voice transcription and emotion marking, structured data verification and completion according to the characteristics of the detection industry, and greatly improves data quality and usability;

[0022] 2. The present application adopts a combination analysis of a multiple linear regression model and a Bayesian probability model, introduces multi-dimensional independent variables such as user industry and historical interaction frequency, accurately extracts industry professional keywords in combination with a TF-IDF algorithm, and significantly improves the accuracy of advisory intention matching degree and confidence determination;

[0023] 3. This invention establishes a three-level progressive response mode: for clear needs, it quickly outputs standardized responses; for ambiguous needs, it provides precise alternative answers and guidance for confirmation; and for highly complex needs, it automatically transfers to human assistants and synchronizes complete auxiliary information, balancing response speed and service accuracy, and reducing user waiting time and repetitive communication.

[0024] 4. This invention enables real-time updates of the knowledge base and dynamic optimization of the model through a data feedback module, continuously adapting to changes in industry policies, projects, and user needs. With long-term use, the service accuracy and adaptability steadily improve. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0026] Figure 2 This is a schematic diagram of the module connection of the present invention;

[0027] Figure 3 This is a data processing flowchart of the present invention;

[0028] Figure 4 This is a diagram of the three-level output response mechanism of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] As attached Figure 2 The AI ​​customer service digital employee system shown is applied to the testing industry and includes a data acquisition module, a data preprocessing module, a data fine processing module, a mathematical model analysis module, a hierarchical output module, a knowledge base module, and a data feedback module. Each module is connected in sequence through a communication link, and the data feedback module is communicatively connected to the hierarchical output module, the knowledge base module, and the data fine processing module.

[0031] The data acquisition module is used to collect raw data from customer service scenarios in the testing industry through preset multi-channel interfaces. The raw data includes testing business consultation text data, user voice interaction recording data, historical customer service dialogue data, structured data from the testing industry knowledge base, and user basic information data. At the same time, the module classifies and marks the collected raw data, including data type, collection time, collection channel, and user identifier, and transmits the marked raw data to the data preprocessing module.

[0032] The data preprocessing module is configured to receive the labeled raw data transmitted by the data collection module, and perform standardization processing on the data according to data types, specifically including text data cleaning and term standardization, speech data transcription and emotion labeling, structured data verification and format unification, to form a standardized preprocessing data set, and transmit the standardized preprocessing data set to the data fine processing module.

[0033] The data fine processing module is configured to receive the standardized preprocessing data set, and perform semantic segmentation and keyword extraction, keyword classification and labeling, and data correlation mapping, to establish a correlation data of “user basic information-consultation text-keyword category” indexed by a user identifier, output a structured feature data set, and transmit the structured feature data set to the mathematical model analysis module.

[0034] The mathematical model analysis module is configured to receive the structured feature data set, and perform quantitative analysis by combining a multiple linear regression model and a Bayesian probability model, calculate a consultation intent matching degree and a posterior probability of a reply content after correction, determine a demand matching result type (clear demand, fuzzy demand, and high fuzzy demand) through a confidence test, and transmit the matching result and the corresponding confidence to the hierarchical output module.

[0035] The hierarchical output module is configured to receive the matching result and the confidence transmitted by the mathematical model analysis module, output a response according to a three-level progressive mode (first-level output: clear demand and fast standardized reply; second-level output: fuzzy demand and accurate answer selection guidance confirmation; and third-level output: high fuzzy demand and artificial transfer auxiliary support), and transmit related data (output level, reply content, user feedback result, and artificial transfer record) of the output response to the data feedback module.

[0036] The knowledge base module is configured to store detection industry professional data and standardized reply content, including detection project name and technical parameters, detection process steps and operation specifications, detection fee pricing standards, detection report index meanings and interpretation rules, detection agency qualification certification information, detection standard regulations, and artificial customer service reply content that has passed the audit, to provide data support for the hierarchical output module.

[0037] The data feedback module is configured to receive the related data of the output response transmitted by the hierarchical output module, audit the artificial customer service reply content, store the content that has passed the audit in the knowledge base module, update the “keyword-reply content” corresponding relationship data, and simultaneously associate the user feedback result and the related data as a new training sample, and transmit the new training sample to the data fine processing module for model parameter adjustment, to realize system closed-loop optimization.

[0038] An AI customer service digital employee method applied to a detection industry, as shown in the accompanying Figure 1 The method comprises the following steps based on the above system implementation:

[0039] As attached Figure 3 The data processing flow shown includes S1, data collection: collecting raw data from customer service scenarios in the testing industry through multiple channel interfaces. The raw data includes testing business consultation text data (including text consultation content sent by users through web pages, APP, mini-program, and SMS, and user voice interaction recording data including voice consultation content sent by users through telephone and voice call functions), historical customer service dialogue data (including all user-customer interaction text and voice-transcribed text data stored in the system within the past 12 months), structured data from the testing industry knowledge base (including data on testing projects, processes, fees, report interpretation, qualifications, regulations, etc.), and user basic information data (including user's industry category, registration time, historical testing business handling records, and consultation frequency). The collected raw data is classified and labeled, including labeling data type, collection time, collection channel, and user identifier, to form labeled raw data.

[0040] S2. Data preprocessing: The labeled raw data is processed through text data processing, speech data processing, and structured data processing to obtain a standardized preprocessed dataset;

[0041] The text data processing refers to cleaning the text data of testing business consultation and historical customer service dialogue, removing special symbols, blank characters, meaningless interjections, and deleting duplicate identical text content; and replacing non-standardized terms with standardized expressions based on a pre-set professional terminology dictionary of the testing industry (which includes commonly used project names, technical terms, qualification names, and regulatory names in the testing industry, in accordance with national and industry standards).

[0042] The voice data processing refers to using voice recognition technology to convert user voice interaction recordings into text data, extracting emotional feature parameters such as speech rate, pitch fluctuation amplitude, volume peak, and pause frequency through voice emotion recognition algorithms, determining the user's emotion type (calm, questioning, anxious, dissatisfied), and marking it to the corresponding speech-to-text data.

[0043] The structured data processing refers to format verification of the structured data in the testing industry knowledge base, checking the completeness of necessary fields such as testing project cost range, process time nodes, report indicator thresholds, and qualification validity period, and supplementing missing fields; unifying data formats to ensure consistent units for numerical data and standardized expression of text data.

[0044] S3. Data Reprocessing: The standardized preprocessed dataset is processed through semantic word segmentation and keyword extraction, keyword classification and labeling, and data association mapping to output a structured feature dataset;

[0045] The semantic word segmentation and keyword extraction method is as follows: adopting a word segmentation algorithm (specifically, a forward maximum matching word segmentation method based on detection industry professional dictionary) to perform semantic word segmentation on the preprocessed text data, and the word segmentation granularity is accurate to the detection industry professional term level (i.e., the minimum word segmentation unit is a professional vocabulary recognized by the detection industry, such as "heavy metal lead detection" and "CMA qualification certification"); the weight value of each word segmentation is calculated by a term frequency-inverse document frequency algorithm (TF-IDF), and the TF-IDF algorithm formula is: TF IDF = TF x IDF, where TF is the ratio of the frequency of the word segmentation appearing in the current text to the total number of words in the current text, and IDF is log 10 (total number of texts / number of texts containing the word segmentation + 1); the word segmentation with a weight value greater than a threshold value (the threshold value is determined according to the word frequency distribution characteristics of the detection industry text, by statistically analyzing the word frequency variance of the historical consultation text data in the past 12 months) is selected as a core keyword (the core keyword is defined as: a detection industry professional term or a key information word that can accurately reflect the core demand of the user consultation).

[0046] The keyword classification labeling method is as follows: based on the detection industry customer service consultation scene classification rules (divided into six categories: detection project consultation, detection process consultation, detection fee consultation, detection report interpretation, detection qualification inquiry, and complaint suggestion, and each category corresponds to a fixed keyword set: the detection project consultation keywords include "detection project", "detection index", "detection parameter", etc.; the detection process consultation keywords include "detection cycle", "detection step", "sampling specification", etc.; the detection fee consultation keywords include "detection fee", "fee standard", "quotation", etc.; the detection report interpretation keywords include "report index", "result interpretation", "qualification standard", etc.; the detection qualification inquiry keywords include "CMA certification", "CNAS qualification", "qualification validity period", etc.; and the complaint suggestion keywords include "complaint", "suggestion", "problem feedback", etc.), the core keywords are classified and labeled to form a keyword-category mapping relationship data.

[0047] The data correlation mapping is as follows: the user basic information data is correlated and mapped with the corresponding consultation text data and keyword classification data to establish a "user basic information-consultation text-keyword category" correlation data indexed by the user identifier (the data includes six fields of user industry category, historical detection business handling record, consultation frequency, consultation text content, core keyword, and keyword category), and an output structured feature data set.

[0048] S4, mathematical model data analysis: based on the structured feature data set, a multiple linear regression model is constructed and calculated, a Bayesian probability model is calculated, and a confidence test is performed to obtain an optimal matching result and a confidence level.

[0049] The multiple linear regression model is constructed and calculated as follows: the keywords category (quantified as a numerical value of 1-6, quantified as 1 for detection project consultation, quantified as 2 for detection process consultation, quantified as 3 for detection fee consultation, quantified as 4 for detection report interpretation, quantified as 5 for detection qualification inquiry, and quantified as 6 for complaint suggestion) in the structured feature data set, the user historical interaction frequency (the total number of consultations in the past 12 months, taking a positive integer value), and the user industry category (quantified as a numerical value of 1-6, quantified as 1 for industrial detection industry, quantified as 2 for food detection industry, quantified as 3 for environmental detection industry, quantified as 4 for pharmaceutical detection industry, quantified as 5 for agricultural product detection industry, and quantified as 6 for other detection related industries) are used as independent variables, and the "consultation intent matching degree" (the value range is 0-1, the numerical value is closer to 1, indicating that the user's consultation intent is more clear, and the matching degree of the demand corresponding to the keywords is higher) is used as the dependent variable, to construct a multiple linear regression model: Y = a0 + a1X1 + a2X2 + a3X3 + ε, wherein a0 is the regression constant term (the value range is 0-0.2), a1, a2, and a3 are the regression coefficients of the keywords category, the user historical interaction frequency, and the user industry category (the value range is 0-0.3), and ε is the error term (subject to a normal distribution with a mean of 0 and a variance of σ2, and σ2 takes a value range of 0.01-0.05); the regression coefficients are calculated and determined by using the least square method based on 100,000 valid interaction data samples with clear intent in the past 12 months (valid interaction data sample is defined as: the user's consultation content is clear, and there is an accurate reply to the interaction data); the objective of the least square method is to minimize the error sum of squares of the model prediction value and the actual intent matching degree; the numerical value of the consultation intent matching degree Y is calculated by substituting the numerical value of the independent variable in the structured feature data set into the model;

[0050] The Bayesian probability model is calculated as follows: based on the corresponding frequency of "core keywords-reply content" in the historical customer service dialogue data, the prior probability of user core keywords (the calculation method is: the total number of times the core keyword appears in the historical data divided by the total number of times all core keywords appear) and the reply content user core keyword (the calculation method is: the number of times the core keyword and the corresponding reply content appear together divided by the total number of times the core keyword appears); the posterior probability is calculated according to the Bayes formula The posterior probability is calculated according to the Bayes formula The reply content user core keyword is weighted and corrected (Y is the consultation intent matching degree calculated by the multiple linear regression model), and the reply content with the highest corrected posterior probability is selected as the optimal matching result of the user's consultation intent;

[0051] The confidence test is as follows: the value of the corrected posterior probability is taken as the confidence (the value range is 0-1, the higher the confidence, the higher the matching result is consistent with the user's true demand), and two confidence thresholds of 95% and 90% are set; when the confidence is ≥95%, it is determined that the demand matching result is clear (clear demand is defined as: the user's consultation intention is clear, the core demand is clear, and the demand can be accurately replied without further confirmation); when 90%≤confidence<95%, it is determined that the demand matching result is fuzzy (fuzzy demand is defined as: the user's consultation intention is basically clear, but there are some ambiguities or information missing, and the demand can be accurately replied after further guidance and confirmation); when the confidence is <90%, it is determined that the demand matching result is high fuzzy (high fuzzy demand is defined as: the user's consultation intention is not clear, involves multi-dimensional complex problems or special scene demand, AI cannot accurately reply independently, and human intervention is required for processing).

[0052] As shown in S5 of FIG. 8, the hierarchical output is based on the confidence and the optimal matching output to output a first-level output, a second-level output and a third-level output. Figure 4

[0053] The first-level output (clear demand response) is as follows: when the confidence is ≥95%, the hierarchical output module directly calls the standardized reply content (standardized reply content is defined as: preset for detecting various clear consultation demands of the industry, the text content conforms to the professional specification, is simple and accurate) corresponding to the optimal matching result in the knowledge base module, and selects the corresponding output form according to the type of the user's original consultation channel: if the user consults through the text channel (web page, APP, small program, SMS), the output is in text form; if the user consults through the voice channel (telephone, voice call function), the text reply content is converted into voice form output, the speech speed of the voice output is 180 words per minute, the tone conforms to the habit of daily communication (the tone fluctuation amplitude is controlled within 5 decibels), and the user can clearly understand;

[0054] The second-level output (fuzzy demand response) is as follows: when 90%≤confidence<95%, the hierarchical output module calls the top three reply contents in the knowledge base module after the correction of the posterior probability as the prepared selected answers, and each prepared selected answer is clearly marked with the corresponding applicable scene (the applicable scene marking content includes the detection item type, the user industry category, the consultation scene, etc. corresponding to the prepared selected answer); at the same time, a guidance confirmation text is generated, which contains specific questions for the fuzzy points (the question content needs to clearly point to the ambiguity or missing information of the user's consultation), and is used to guide the user to further clarify the consultation demand; then the prepared selected answers and the guidance confirmation text are output through the user's original consultation channel in the order of "prepared selected answers in front and guidance text behind".

[0055] ​The third level output (high fuzzy demand response) is as follows: when the confidence is < 90%, the hierarchical output module automatically triggers the manual customer service transfer instruction, extracts the corresponding "user basic information-consultation text-keyword category" associated data of the user, the matching result and the confidence obtained by mathematical model analysis, and forms an auxiliary information package (the auxiliary information package is defined as: a comprehensive information set that can help manual customer service quickly understand the user's situation and accurately position the user's demand); the auxiliary information package is transmitted to the manual customer service terminal through the internal communication link, and the manual customer service terminal automatically pops up the auxiliary information package after receiving it for manual customer service to quickly check; at the same time, the hierarchical output module sends a transfer prompt information to the user, which clearly informs the user that the current manual customer service has been transferred and the estimated waiting time, and the estimated waiting time is calculated as: estimated waiting time = number of people in line × average processing time / number of online manual customer service personnel, wherein the average processing time is the average time of manual customer service processing a single consultation in the past 1 hour (calculated as: the ratio of the total time of manual customer service processing consultations in the past 1 hour to the total number of consultations processed).

[0056] S6, closed loop optimization: optimizing the output results, including knowledge base updating and model optimization;

[0057] The knowledge base updating is as follows: the data feedback module audits the reply content generated by manual customer service in handling high fuzzy demand, and the auditing standard includes: whether the reply content conforms to the detection industry professional specification (whether it follows the detection industry national standard, industry standard and internal specification of the institution), whether it accurately answers the user's question (whether it gives a clear answer to the core demand of the user's consultation), and whether it contains necessary operation guide (if the user's demand involves subsequent operation, whether the reply content contains specific operation steps or contact information); after the audit is passed, the reply content is associated with the corresponding user consultation text, core keyword and keyword category, and stored in the knowledge base module to update the "keyword-reply content" corresponding relationship data in the knowledge base module.

[0058] The model optimization is as follows: the data feedback module associates the feedback results of the user on the first level output and the second level output (the feedback results include "satisfied", "dissatisfied" and supplementary consultation content, "satisfied" means that the user approves the reply content, "dissatisfied" means that the user thinks that the reply content does not solve his demand, and the supplementary consultation content is the user's further question for the original consultation) with the corresponding structured feature data and model analysis result as new training samples, and transmits them to the data processing module; the data processing module adjusts the keyword extraction preset threshold (the adjustment range is ±0.02, which is determined according to the keyword weight value distribution characteristics of the new training samples) and the regression coefficient of the multiple linear regression model (the adjustment range is ±0.01, which is determined according to the model prediction error of the new training samples) every 24 hours, to continuously optimize the model analysis accuracy.

[0059] The embodiment takes the food detection industry (corresponding to X3=2) as the application scenario, and the specific implementation process is as follows:

[0060] Send the text consultation "What is the detection period of heavy metal lead in food?" through the APP end.

[0061] Data collection: The data collection module collects the consultation text "What is the detection period of heavy metal lead in food?", the user A basic information (registered time 2023 January, historical consultation frequency 3 times, corresponding to X2=3, no historical detection business record), the knowledge base related structured data of "food heavy metal lead detection" (including detection project technical parameters, process steps, period 3 working days, fee standard, etc.), and marks the data type as "detection business consultation text data", "user basic information data" and "knowledge base structured data", the collection time is the current system time, the collection channel is "APP end", the user identifier is "user A", and the original data after marking is formed.

[0062] Data preprocessing: The data preprocessing module cleans the consultation text, and there are no special symbols, blank characters, meaningless mood adverbs, and no repeated content, and the original text is directly retained; based on the detection industry professional term dictionary, "food", "heavy metal lead" and "detection period" are all standardized terms and do not need to be replaced; the knowledge base data is verified to confirm that the period field of "food heavy metal lead detection" is "3 working days", there is no missing, the data format is unified (the time unit is "working day"), and the standardized preprocessing data set is formed.

[0063] Data reprocessing: The data processing module adopts the forward maximum matching segmentation method based on the detection industry professional dictionary, and the consultation text is segmented into "food", "heavy metal lead" and "detection period"; the weight value is calculated by the TF-IDF algorithm, the TF value of "food" is 1 / 3, the IDF value is log 10 (100000 / 25000+1)=log 10 (5)=0.699, the weight value is approximately 0.233; the TF value of "heavy metal lead" is 1 / 3, the IDF value is log 10 (100000 / 5000+1)=log10(21)=1.322, the weight value is approximately 0.441; the TF value of "detection period" is 1 / 3, the IDF value is log 10 (100000 / 18000+1)=log 10(6.555)≈0.816, weight value ≈0.272; the sum of the total weight values is 0.946, and each word weight value is greater than 0.6, which is determined as a core keyword; according to the classification rule, “detection cycle” belongs to the detection process consultation class, and the three core keywords are uniformly marked as “detection process consultation class” (corresponding to X1=2); taking the user identifier “user A” as an index, the user basic information (food detection industry, historical consultation 3 times), consultation text, keyword category are associated to form a structured feature data set.

[0064] Mathematical model analysis: the mathematical model analysis module substitutes X1=2, X2=3, X3=2 into the multiple linear regression model (preset a0=0.1, a1=0.2, a2=0.1, a3=0.15, ε=0.02), and calculates Y=0.1+0.2×2+0.1×3+0.15×2+0.02=0.1+0.4+0.3+0.3+0.02=1.12, because the value range of Y is 0-1, take the upper limit value 1.0; based on historical data, the core keyword “food + heavy metal lead + detection cycle” and the reply content “the detection cycle of heavy metal lead in food is 3 working days” appear together 1200 times, the total number of keyword combinations is 1224 times, and the reply content is the user core keyword; the user core keyword; the reply content; according to Bayes formula, the posterior probability P(A|B)=(0.980×0.0024) / 0.003≈0.784 is calculated; through the correction formula, the confidence is 98.0%≥95%, and it is judged as a clear demand matching result.

[0065] Hierarchical output: the hierarchical output module determines that it is a clear demand, and retrieves the corresponding standardized reply “the detection cycle of heavy metal lead in food is 3 working days” in the knowledge base; because the user consultation channel is APP (text channel), the text is output to user A through APP.

[0066] Closed loop optimization: user A clicks the “satisfied” feedback button, and the data feedback module stores the interaction data (user consultation text, core keyword, reply content, and feedback result “satisfied”) in association, as a new training sample, and transmits it to the data processing module for model parameter adjustment in the next 24 hours.

[0067] The embodiment takes the environmental detection industry (corresponding to X3=3) as an application scenario, and the specific implementation process is as follows:

[0068] Send voice consultation “how much does factory wastewater detection cost” through phone;

[0069] Data Collection: The data collection module collects voice recording data, user B's basic information (registered in May 2022, historical consultation frequency of 5 times, corresponding to X2=5, no historical detection business records), and "factory wastewater detection" cost-related data in the knowledge base (regular project 800-1500 yuan, full item detection 2000-3000 yuan, specific pollutant detection 500-1000 yuan / item). The data types are labeled as "user voice interaction recording data", "user basic information data", and "knowledge base structured data". The collection time is the current system time, the collection channel is "telephone terminal", and the user identifier is "user B". The original data is formed after labeling.

[0070] Data Preprocessing: The data preprocessing module uses speech recognition technology to transcribe the voice recording into text "How much does factory wastewater detection cost?"; Extract emotion feature parameters: speech rate 170 words / minute, pitch fluctuation amplitude 3 decibels, volume peak 60 decibels, pause frequency 2 times / minute, determine emotion type as calm, and mark to the transcribed text; Clean the text to remove adverbs "need" and "how much", and keep the core content "factory wastewater detection cost"; Based on the professional term dictionary, "factory wastewater detection" and "cost" are standardized expressions; Check the knowledge base cost data, the field is complete, and the unit is unified to "yuan", forming a standardized preprocessing data set.

[0071] Data Re-processing: The data processing module uses the forward maximum matching segmentation method to segment the transcribed text into "factory wastewater detection" and "cost"; Calculate the weight value by TF-IDF algorithm, "factory wastewater detection" TF value is 1 / 2, IDF value is log 10 (100000 / 8000+1)=log 10 (13.5)=1.130, weight value ≈ 0.565; "cost" TF value is 1 / 2, IDF value is log 10 (100000 / 30000+1)=log 10 (4.333)≈0.637, weight value ≈ 0.318; Total weight value 0.883, all greater than 0.6, determined as core keywords; According to the classification rules, "cost" belongs to the detection cost consultation category, and the two core keywords are labeled as "detection cost consultation category" (corresponding to X1=3); Take the user identifier "user B" as the index, associate user basic information (environmental detection industry, historical consultation 5 times), consultation text, and keyword category, form a structured feature data set.

[0072] Mathematical Model Analysis: Substitute X1=3, X2=5, X3=3 into the multiple linear regression model (a0=0.1, a1=0.2, a2=0.1, a3=0.15, ε=0.03), calculate:

[0073] Y = 0.1 + 0.2 x 3 + 0.1 x 5 + 0.15 x 3 + 0.03 = 0.1 + 0.6 + 0.5 + 0.45 + 0.03 = 1.68, normalized to 0.96 (normalization method: normalization, Y max The maximum historical calculation value is 2.0); based on historical data, the posterior probabilities of the three types of replies are calculated: the reply keyword "Factory wastewater routine detection project cost 800-1500 yuan", the reply keyword "Full item detection 2000-3000 yuan", and the reply keyword "Specific pollutant detection 500-1000 yuan / item"; the corrected probabilities are 0.960 x 0.96 = 0.9216, 0.947 x 0.96 = 0.9091, and 0.923 x 0.96 = 0.8861, respectively; the highest confidence is 92.16%, which is in the interval of 90% < confidence < 95%, and is determined as a fuzzy demand matching result.

[0074] Hierarchical output: The hierarchical output module retrieves the top three alternative answers and labels the applicable scenarios respectively: "1. Factory wastewater routine detection project: cost 800-1500 yuan (applicable scenario: industrial enterprise daily wastewater discharge compliance detection, covering basic indicators such as pH value, chemical oxygen demand, five-day biochemical oxygen demand, etc.); 2. Factory wastewater full item detection project: cost 2000-3000 yuan (applicable scenario: enterprise wastewater treatment process optimization, environmental protection acceptance detection, covering regular indicators and special indicators such as heavy metals and volatile organic compounds); 3. Factory wastewater specific pollutant detection: cost 500-1000 yuan / item (applicable scenario: detection of a single or multiple pollutants specified by the user, such as lead, mercury, formaldehyde, etc.)"; generate a guided confirmation text "Do you need to detect a regular item, a full item, or a specific pollutant item?"; because the user's consultation channel is the phone end (voice channel), the alternative answers and the guide text are converted into voice (speech rate 180 words / minute, tone fluctuation amplitude 4 decibels), and output through the phone end.

[0075] Closed-loop optimization: User B replies "regular item" in voice, the data collection module collects this supplementary consultation data, repeats the above preprocessing and fine processing steps, and reextracts the core keywords "factory wastewater detection" "regular item" "cost", which are classified as "detection cost consultation type"; the mathematical model analyzes and calculates the confidence of 98.5%, triggers the first-level output, and replies in voice "The cost of factory wastewater routine detection project is 800-1500 yuan, the detection period is 4 working days, do I need to help you make an appointment for detection service?" User B feedback "satisfied", the data feedback module associates the supplementary consultation content, the final reply, and the feedback result as new training samples and transmits them to the data fine processing module.

[0076] Secondly: the embodiment of the present application discloses only the structure related to the embodiment of the present application, other structures can refer to the general design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0077] Finally: the above only for the preferred embodiment of the present application, and does not limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for using AI-powered digital customer service staff in the testing industry, characterized in that, include: S1. Data Collection: Collect raw data from customer service scenarios in the testing industry through multiple channels and interfaces. Classify and label the collected raw data, including labeling the data type, collection time, collection channel, and user identifier, to form labeled raw data. S2. Data preprocessing: The labeled raw data is processed through text data processing, speech data processing, and structured data processing to obtain a standardized preprocessed dataset; S3. Data Reprocessing: The standardized preprocessed dataset is processed through semantic word segmentation and keyword extraction, keyword classification and labeling, and data association mapping to output a structured feature dataset; S4. Mathematical Model Data Analysis: Based on the structured feature dataset, construct and calculate the multiple linear regression model, calculate the Bayesian probability model, and perform confidence tests to obtain the optimal matching results and confidence scores. S5. Tiered Output: Based on confidence level and optimal matching, outputs first-level, second-level, and third-level outputs; S6. Closed-loop optimization: Optimize the output results, including knowledge base updates and model optimization.

2. The AI ​​customer service digital employee method applied to the testing industry according to claim 1, characterized in that: The raw data includes testing business consultation text data, historical customer service dialogue data, structured data from the testing industry knowledge base, and basic user information data; The text data for testing business inquiries includes text inquiries sent by users through web pages, apps, mini programs, and SMS; the voice interaction recording data includes voice inquiries sent by users through telephones and voice call functions. The historical customer service dialogue data includes all user-customer service interaction texts and speech-to-text data stored in the system over the past 12 months. The structured data in the testing industry knowledge base includes testing items, processes, costs, report interpretation, qualifications, and regulatory data. The user's basic information data includes the user's industry category, registration time, historical testing business processing records, and consultation frequency.

3. The AI ​​customer service digital employee method for the testing industry according to claim 1, characterized in that: The text data processing refers to cleaning the text data of testing business consultations and historical customer service dialogues, removing special symbols, whitespace characters, meaningless interjections, and deleting duplicate identical text content; and replacing non-standard terms with standardized expressions based on a pre-set professional terminology dictionary for the testing industry. The voice data processing refers to using voice recognition technology to convert user voice interaction recording data into text data, extracting emotional feature parameters through voice emotion recognition algorithms, determining the user's emotion type (including calm, questioning, anxious, and dissatisfied), and marking it to the corresponding speech-to-text data. The structured data processing refers to performing format validation on the structured data in the testing industry knowledge base, checking the completeness of testing fields, supplementing missing fields, and standardizing data formats.

4. The AI ​​customer service digital employee method applied to the testing industry according to claim 1, characterized in that: The semantic word segmentation and keyword extraction method is as follows: the preprocessed text data is semantically segmented using a word segmentation algorithm, with the segmentation granularity accurate to the level of detecting industry professional terms. The weight value of each word is calculated using the term frequency-inverse document frequency algorithm, and words with weight values ​​greater than the threshold are selected as core keywords. The keyword classification and labeling method is as follows: Based on the classification rules of customer service consultation scenarios in the testing industry, core keywords are classified and labeled to form keyword-category mapping relationship data; The data association mapping is as follows: the user basic information data is associated with the corresponding consultation text data and keyword classification data, and a "user basic information-consultation text-keyword category" association data indexed by the user identifier is established, and a structured feature dataset is output.

5. The AI ​​customer service digital employee method for the testing industry according to claim 1, characterized in that: The construction and calculation of the multiple linear regression model are as follows: using keyword categories, user historical interaction frequency, and user industry category in the structured feature dataset as independent variables, and "consultation intent matching degree" as the dependent variable, a multiple linear regression model is constructed; the values ​​of the independent variables in the structured feature dataset are substituted into the model to calculate the specific value of consultation intent matching degree Y. The Bayesian probability model is calculated as follows: Based on the corresponding frequency of "core keywords - reply content" in historical customer service dialogue data, the prior probability of user core keywords and reply content user core keywords are determined. The posterior probability is calculated according to the Bayesian formula. The posterior probability is weighted and corrected by the reply content user core keywords through the correction formula. The reply content with the highest corrected posterior probability is selected as the optimal matching result of the user's consultation intent. The confidence level test is as follows: the value of the corrected posterior probability is used as the confidence level, and two confidence level thresholds of 95% and 90% are set; when the confidence level is ≥95%, it is determined to be a clear demand matching result; when 90%≤confidence level<95%, it is determined to be a fuzzy demand matching result; when the confidence level<90%, it is determined to be a highly fuzzy demand matching result.

6. The AI ​​customer service digital employee method for the testing industry according to claim 1, characterized in that: The first-level output is as follows: When the confidence level is ≥95%, the hierarchical output module directly retrieves the standardized response content corresponding to the optimal matching result from the knowledge base module, and selects the corresponding output format according to the type of the user's original consultation channel: if the user consults through a text channel, the output is in text form; if the user consults through a voice channel, the text response content is converted into voice form for output. The secondary output is as follows: When 90% ≤ confidence level < 95%, the hierarchical output module retrieves the top three responses from the knowledge base module with the corrected posterior probability as precise candidate answers, each candidate answer clearly marked with its corresponding applicable scenario; simultaneously, a guidance confirmation text is generated, which contains specific questions addressing points of ambiguity, used to guide the user to further clarify their consultation needs; subsequently, the precise candidate answers and the guidance confirmation text are output through the user's original consultation channel in the order of "candidate answers first, guidance text second"; The three-level output is as follows: When the confidence level is <90%, the hierarchical output module automatically triggers a human customer service transfer instruction, and simultaneously extracts the user's corresponding "user basic information-consultation text-keyword category" association data, the matching results obtained from mathematical model analysis, and the confidence level to form an auxiliary information package; the auxiliary information package is transmitted to the human customer service terminal through an internal communication link, and the human customer service terminal automatically displays the auxiliary information package in a pop-up window after receiving it, allowing human customer service to quickly review it; at the same time, the hierarchical output module sends a transfer prompt message to the user, clearly informing the user that they have been transferred to human customer service and the estimated waiting time.

7. The AI ​​customer service digital employee method applied to the testing industry according to claim 1, characterized in that: The knowledge base is updated as follows: The data feedback module reviews the response content generated by the human customer service in handling highly ambiguous requests. After the review is passed, the response content is associated with the corresponding user inquiry text, core keywords, and keyword categories, and stored in the knowledge base module. The "keyword-response content" correspondence data in the knowledge base module is updated. The model is optimized as follows: The data feedback module associates the user's feedback results on the first-level output and the second-level output with the corresponding structured feature data and model analysis results, and transmits them as new training samples to the data refinement module; the data refinement module adjusts the preset threshold for keyword extraction and the regression coefficients of the multiple linear regression model every 24 hours to continuously optimize the model analysis accuracy.

8. An AI customer service digital employee system for the testing industry, used to implement the AI ​​customer service digital employee method for the testing industry as described in any one of claims 1-7, characterized in that, include: The system includes a data acquisition module, a data preprocessing module, a data fine processing module, a mathematical model analysis module, a hierarchical output module, a knowledge base module, and a data feedback module. The data acquisition module is used to collect raw data from customer service scenarios in the testing industry through preset multi-channel interfaces. The raw data includes testing business consultation text data, user voice interaction recording data, historical customer service dialogue data, structured data from the testing industry knowledge base, and user basic information data. At the same time, the module classifies and marks the collected raw data, including data type, collection time, collection channel, and user identifier, and transmits the marked raw data to the data preprocessing module. The data preprocessing module is used to receive the labeled raw data transmitted by the data acquisition module, and perform standardization processing according to the data type. Specifically, it includes text data cleaning and terminology standardization, speech data transcription and emotion tagging, structured data verification and format unification to form a standardized preprocessed dataset, which is then transmitted to the data fine processing module. The data fine processing module is used to receive a standardized preprocessed dataset, perform semantic word segmentation and keyword extraction, keyword classification and labeling, data association mapping, establish "user basic information-consultation text-keyword category" association data indexed by user identifier, output a structured feature dataset, and transmit it to the mathematical model analysis module. The mathematical model analysis module is used to receive structured feature datasets, perform quantitative analysis using a combination of multiple linear regression and Bayesian probability models, calculate the matching degree of consultation intent and the posterior probability after correcting the response content, determine the type of demand matching result through confidence test, and transmit the matching result and the corresponding confidence level to the hierarchical output module. The hierarchical output module is used to receive the matching results and confidence levels transmitted by the mathematical model analysis module, output the response according to the three-level progressive mode, and transmit the relevant data of the output response to the data feedback module. The knowledge base module is used to store professional data and standardized responses in the testing industry, including the name and technical parameters of testing items, testing process steps and operating procedures, testing fee pricing standards, the meaning and interpretation rules of various indicators in the testing report, the qualification certification information of testing institutions, testing standard and regulatory provisions, and the content of human customer service responses that have been reviewed and approved, providing data support for the hierarchical output module. The data feedback module is used to receive output response data transmitted from the hierarchical output module, review the content of human customer service replies, store the approved content in the knowledge base module, and update the "keyword-reply content" correspondence data. At the same time, it associates user feedback results with relevant data as new training samples and transmits them to the data fine processing module for model parameter adjustment to achieve closed-loop optimization of the system.

Images