A knowledge base optimization method based on customer service dialogue quality inspection and related equipment

Abstract

This application provides a knowledge base optimization method and related equipment based on customer service dialogue quality inspection. The method involves acquiring customer service dialogue data and a customer service knowledge base; performing quality inspection on the customer service dialogue data; filtering target dialogue data from the customer service dialogue data based on the quality inspection results; extracting multiple candidate question-answer pairs from the target dialogue data; grouping the multiple candidate question-answer pairs based on the semantic relationships between the multiple candidate questions to obtain question-answer groups; determining the core representative question from each question-answer group; and updating the customer service knowledge base based on the core representative question and its corresponding candidate answers. By filtering target dialogues based on the quality inspection results, extracting candidate question-answer pairs and grouping them according to semantic relationships, extracting core representative questions to update the customer service knowledge base, and then back-verifying the quality inspection rules after the update, a closed-loop feedback is formed, ultimately achieving continuous collaborative optimization of the customer service knowledge base and quality inspection rules.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a knowledge base optimization method and related equipment based on customer service dialogue quality inspection. Background Technology

[0002] With the continuous development of large language models and intelligent customer service technologies, customer service systems are gradually gaining the ability to automatically respond to user inquiries. The completeness and accuracy of the knowledge base directly impact the service effectiveness of intelligent customer service. Customer service dialogues contain a large amount of business processing flow and question-and-answer information. Automated quality inspection, knowledge extraction, and continuous optimization of customer service dialogues have become important technical directions for improving the service capabilities of intelligent customer service and the quality of the knowledge base.

[0003] Existing technologies typically perform semantic analysis and intent recognition on customer service conversations, and combine this with a pre-set knowledge base to complete the conversation quality score; or they collect business documents and conversation data, extract knowledge items using a large language model, compare them with the knowledge base for consistency, and then update the knowledge base through version management. However, because customer service conversation quality inspection and knowledge base updates are independent in existing technologies, the quality inspection results cannot be used to guide knowledge accumulation, and the actual application effect of the knowledge base after its launch cannot be used to optimize the quality inspection standards. This results in a lack of a closed-loop feedback mechanism between customer service quality inspection, knowledge extraction, and knowledge base optimization, making it difficult to achieve continuous collaborative optimization of the knowledge base and quality inspection rules. Summary of the Invention

[0004] In view of the shortcomings of the prior art, this application provides a knowledge base optimization method and related equipment based on customer service dialogue quality inspection, so as to solve the problem of difficulty in achieving continuous collaborative optimization of knowledge base and quality inspection rules.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] This application provides a knowledge base optimization method and related equipment based on customer service dialogue quality inspection. The method involves acquiring customer service dialogue data and a customer service knowledge base; performing quality inspection on the customer service dialogue data; filtering target dialogue data from the customer service dialogue data based on the quality inspection results; extracting multiple candidate question-answer pairs from the target dialogue data; grouping the multiple candidate question-answer pairs based on the semantic relationships between the multiple candidate questions to obtain question-answer groups; determining the core representative question from each question-answer group; and updating the customer service knowledge base based on the core representative question and its corresponding candidate answers. By filtering target dialogues based on the quality inspection results, extracting candidate question-answer pairs and grouping them according to semantic relationships, extracting core representative questions to update the customer service knowledge base, and then back-verifying the quality inspection rules after the update, a closed-loop feedback is formed, ultimately achieving continuous collaborative optimization of the customer service knowledge base and quality inspection rules. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0008] Figure 1 A flowchart illustrating a knowledge base optimization method based on customer service dialogue quality inspection provided in this application embodiment;

[0009] Figure 2 A flowchart of another knowledge base optimization method based on customer service dialogue quality inspection provided in this application embodiment;

[0010] Figure 3 A flowchart of another knowledge base optimization method based on customer service dialogue quality inspection provided in this application embodiment;

[0011] Figure 4 A schematic diagram of the architecture of a knowledge base optimization device based on customer service dialogue quality inspection provided in this application embodiment;

[0012] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0015] like Figure 1The flowchart shown is a knowledge base optimization method based on customer service dialogue quality inspection provided in this application embodiment, which specifically includes the following steps:

[0016] S101: Obtain customer service dialogue data and customer service knowledge base.

[0017] Among them, customer service dialogue data is structured data formed after standardizing the customer service communication process, and customer service knowledge base is used to store knowledge items that can be accessed by intelligent customer service.

[0018] It should be noted that customer service dialogue data can come from instant messaging online conversations and national service hotline voice conversations. Instant messaging online conversations are text, image, or emoticon message exchanges initiated by users through the company's online chat channels; national service hotline voice conversations are conversations formed by users interacting with human agents in real time via telephone lines.

[0019] Optionally, in another embodiment of this application, the specific implementation of obtaining customer service dialogue data includes processes A1 to A8.

[0020] A1: Obtain the original customer service dialogue data.

[0021] Specifically, raw customer service dialogue data can be read from instant messaging systems, call center systems, or dialogue log storage systems within a preset time range, and a dialogue identifier can be assigned to each raw customer service dialogue data.

[0022] Optionally, the original customer service dialogue data can also be associated with customer service identifiers, user identifiers, service platforms, start time, end time, business channels, and work order identifiers.

[0023] A2: Determine the data type corresponding to the original customer service dialogue data. Data types include instant messaging text type and voice type.

[0024] Specifically, the data type can be determined based on the file format, data source identifier, or session medium field. JSON logs, message queue records, or chat logs can be identified as instant messaging text; audio files or voice streams can be identified as voice.

[0025] A3: If the data type is voice, then the original customer service dialogue data will be converted into voice to obtain the original transcribed text.

[0026] It should be noted that speech conversion includes environmental noise reduction, voice activity detection, and speech recognition. Environmental noise reduction is performed on the original customer service dialogue data, and audio segments containing speech content are identified. These audio segments are then input into an automatic speech recognition model to obtain the original transcribed text.

[0027] Optionally, the speech conversion result can be associated with the audio timeline so that each speech item in the original transcribed text retains its corresponding start and end time.

[0028] A4: If the data type is instant messaging text, then extract the original dialogue text from the original customer service dialogue data.

[0029] Specifically, the original log file is parsed to extract timestamps, senders, and conversation content, and system notifications, empty messages, duplicate messages, and unparseable attachment placeholder content are filtered out.

[0030] A5: Extract multiple statements from the original transcribed text or original dialogue text.

[0031] The content of the speech refers to the text content expressed by the same speaker within the corresponding time period.

[0032] Specifically, text can be segmented into sentences based on pause duration, message interval, punctuation boundaries, or changes in the sender, resulting in multiple messages.

[0033] A6: Based on the dialogue context corresponding to each statement, determine the speaking role corresponding to each statement.

[0034] The speaking roles include customer service representatives and users. The dialogue context includes the preceding and following statements, business terminology, question-and-answer relationships, address methods, expression of operational instructions, and expression of requests.

[0035] Specifically, multiple statements and their chronological order can be input into a preset joint instruction template, and the large language model will determine the speaker's role based on contextual reference relationships and interactive speech features.

[0036] Optionally, when the original log already contains the sender identifier, the sender identifier can be used as a constraint to determine the speaker role, and semantic inference can only be performed on speech content that lacks a sender identifier or has an abnormal sender identifier.

[0037] A7: Add the role identifier corresponding to the speaking role to the corresponding speaking content to obtain the role label text.

[0038] For example, the role identifier can be customer service or user, and the role label text can be organized in the form of "timestamp + role identifier + speech content".

[0039] A8: Standardize the character annotation text to obtain standardized text.

[0040] It should be noted that standardization processing can include typo correction, homophone correction, elimination of redundant colloquial words, standardization of industry terminology, resolution of pronoun references, and processing of semantic coherence.

[0041] Specifically, the standardization process preserves the original business intent and factual information, removes words such as "um," "ah," and "that" that do not carry business semantics, and unifies different expressions of the same business object into preset terms.

[0042] In addition, the correspondence between the original text and the standardized text can be saved, and the processing confidence level can be recorded so that the standardization results can be backtracked.

[0043] A9: Combine the standardized texts according to their corresponding time sequence to obtain customer service dialogue data.

[0044] Specifically, customer service dialogue data can be stored using structured data objects. A structured data object includes at least a dialogue identifier, timestamp, role identifier, standardized text, raw text, and processing confidence level. Its specific format is: {Dialogue ID, Timestamp, Role Identifier [Customer Service / User], Standardized Text, Raw Text, Processing Confidence Level}.

[0045] Optionally, after the combination is completed, the logic of role alternation, text integrity, and chronological order are verified, and rule backtracking correction is performed on abnormal segments. When multiple consecutive statements are marked as the same speaker role and the semantic relationship does not meet the condition of continuous expression, rules or a large language model can be invoked to redetermine the speaker role.

[0046] It should be noted that the customer service knowledge base can include business questions, standard answers, knowledge categories, knowledge source identifiers, knowledge status, version information, and release date. The knowledge source identifier is used to distinguish between manually maintained knowledge entries, business document knowledge entries, and knowledge entries extracted from customer service dialogue data.

[0047] For example, customer service conversation data in the form of text messages is as follows:

[0048] Dialogue ID: xxx;

[0049] Customer Service: xx;

[0050] Customer service: Hello, you have been transferred to a live agent. It's a pleasure to assist you.

[0051] User: Hello;

[0052] Customer service: Hello, what kind of problem do you have?

[0053] User: I still have 4 unused credit cards today, please request a refund;

[0054] Customer service: Is this a prepaid card from a certain business platform?

[0055] User: Yes;

[0056] Customer service: Could you please send me the phone number displayed in the upper left corner of the mini-program?

[0057] User: 12345678910;

[0058] Customer service: Please wait a moment;

[0059] Customer service: A refund has been requested;

[0060] Customer service: If you need a prepaid card next time, you can process the refund yourself on this page;

[0061] Customer service: Page link.

[0062] For example, customer service dialogue data in the form of voice messages is as follows:

[0063] Dialogue ID: xxx;

[0064] Customer Service: xx;

[0065] Call time: [Date and Time]

[0066] Customer Service: Hello, this is employee number 12345. How can I help you?

[0067] User: I have a package order placed on your platform by a certain carrier. Can I provide the tracking number first?

[0068] Customer service: Okay, please provide the order number.

[0069] User: The one with the last four digits of the serial number being 137 kg. I'd like to confirm whether this package is for delivery or self-pickup? The delivery station hasn't contacted me, but the tracking number shows it has arrived.

[0070] Customer Service: Regarding delivery methods, if you did not specifically select "self-pickup" or "door-to-door delivery" when placing your order, the default rules usually depend on the specific carrier's standards. Did you place your order directly on the platform?

[0071] User: Yes, I placed an order on your platform. Since I didn't choose self-pickup, shouldn't it be delivery? If it's a self-pickup service that requires an extra fee, why wasn't this clearly stated on the order page? If that's the case, I wouldn't choose this carrier and would choose another express delivery company instead.

[0072] Customer Service: We understand your concerns. Service standards vary slightly between carriers. For example, some courier services typically offer door-to-door delivery by default, but some express delivery products may require confirmation. Currently, your main problem is that the branch office hasn't contacted you and the delivery status is abnormal, is that correct?

[0073] User: That's right. First, I need them to deliver immediately; second, there's a problem with the tracking information. Last Friday at 9:33 AM it showed "delivering" with a delivery person's phone number, but by noon it showed "customer requested refusal," without any reason given. Last night it showed a refusal again. The delivery person never called me, and I didn't request a refusal. The goods are urgent; they neither contact me nor deliver, just mark it as refused. What kind of operation is this?

[0074] Customer Service: We sincerely apologize for the inconvenience. Regarding this situation, I will immediately file a complaint with the carrier to verify whether delivery is possible and arrange delivery as soon as possible. I will also investigate the reasons for the false signature / refusal to accept delivery.

[0075] User: Your service is terrible.

[0076] Customer Service: We are truly sorry; this was our oversight. I will expedite the process with the carrier and continue to follow up until the issue is resolved. Is there anything else we can assist you with?

[0077] User: No, please process it quickly.

[0078] S102: Perform quality inspection on customer service dialogue data and obtain the quality inspection results.

[0079] The quality inspection results are used to characterize the service quality corresponding to the customer service dialogue data. They may include dialogue category, dimension score corresponding to each target quality inspection indicator, comprehensive score, problem type, rectification suggestions, and quality inspection tags.

[0080] Optionally, in another embodiment of this application, the specific implementation of step S102 includes processes B1 to B3.

[0081] B1: Determine the dialogue category corresponding to the customer service dialogue data.

[0082] The dialogue categories include business consultation and standard process categories.

[0083] The business consultation category is used to represent customer service dialogues that mainly consist of knowledge-based Q&A, while the standard process category is used to represent customer service dialogues that need to be executed according to preset processing steps (i.e., dialogues that are high-frequency, repetitive, multi-step, and require the use of tools are identified as standard process categories).

[0084] Optionally, the rule intent corresponding to the standard process category may include expediting allocation, expediting pickup, expediting logistics, expediting delivery, payment of fees, weight disputes, fee disputes, wrong items, missing items, damage, loss, order cancellation, modification of order information, cancellation of complaints, completion of work orders, member refunds, prepaid card refunds, issuance of invoices, or interception of returns.

[0085] The rules and intents corresponding to the business consultation categories can include account usage, printing issues, electronic waybills, electronic invoices, printer-related matters, other settings, template settings, e-commerce synchronization, member management, import and export, logistics tracking, SMS service, client-side order tracking, contracted clients, webpage tagging, wallet management, employee management, blacklisting numbers, function usage, and interface usage.

[0086] Optionally, in another embodiment of this application, the specific implementation of process B1 includes processes C1 to C4.

[0087] C1: Perform intent recognition on customer service dialogue data to obtain the dialogue intent.

[0088] Specifically, standardized text can be input into a large language model, which then extracts user requests, business objects, target operations, and processing status based on feature vectors from a classification library, and forms dialogue intent based on the extraction results.

[0089] C2: Extract features from the dialogue intent to obtain dialogue intent features.

[0090] Specifically, pre-trained text embedding models can be used to convert dialogue intent into fixed-dimensional feature vectors.

[0091] C3: Obtain the rule intent features corresponding to multiple rule intents in the rule base.

[0092] It should be noted that the rule base includes a business consultation rule base and a standard process rule base. Each rule intent has been pre-vectorized and its characteristics have been saved.

[0093] C4: Calculate the similarity between the dialogue intent features and the rule intent features to obtain the intent matching result.

[0094] For example, cosine similarity can be used to calculate the similarity between dialogue intent features and each rule intent feature, and the rule intent with the highest similarity and its corresponding similarity can be used as the intent matching result.

[0095] C5: Determine the dialogue category corresponding to the customer service dialogue data based on the intent matching results.

[0096] Specifically, when the rule intent with the highest similarity belongs to the business consultation rule base and the similarity reaches the preset intent threshold, the dialogue category is determined to be the business consultation category; when the rule intent with the highest similarity belongs to the standard process rule base and the similarity reaches the preset intent threshold, the dialogue category is determined to be the standard process category.

[0097] Optionally, when the highest similarity does not reach the preset intent threshold, the dialogue category can be marked as a category to be confirmed, or the dialogue category can be determined through business rules and manual review.

[0098] In addition, when outputting the dialogue category corresponding to the customer service dialogue data, the corresponding confidence score will also be output.

[0099] B2: Based on the dialogue category, determine the target quality inspection indicator corresponding to the dialogue category from multiple quality inspection indicators.

[0100] Several quality inspection indicators can be configured according to six dimensions: service attitude, business professionalism, communication skills, problem-solving, service efficiency, and risk compliance. See Table 1 below for details of each quality inspection indicator.

[0101] Table 1

[0102]

[0103] Optionally, the target quality inspection indicators for the business consulting category focus on business expertise and problem-solving, while the target quality inspection indicators for the standard process category focus on problem-solving and risk compliance.

[0104] B3: Based on the target quality inspection indicators, the customer service dialogue data is scored to obtain the quality inspection results.

[0105] Specifically, the large language model extracts the corresponding indicator values ​​for each target quality inspection indicator based on the rule base and outputs the indicator values ​​in a structured format. The scoring module determines the score for each dimension based on the score range of the indicator value and the preset calculation rules.

[0106] For example, the service attitude dimension can be calculated by adding the lowest score in the corresponding score range of the indicator to the coverage rate of positive emotion words; the business professionalism dimension can be calculated by adding the lowest score in the corresponding score range of the indicator to the standardization rate of processing flow execution; and the communication expression dimension can be calculated by adding the lowest score in the corresponding score range of the indicator to the semantic coherence score.

[0107] The problem-solving dimension can be calculated by multiplying the lowest score in the corresponding score range of the indicator with the number of alternative solutions and a preset coefficient (e.g., 0.2); the service efficiency dimension can be calculated by taking the highest score in the corresponding score range; and the risk compliance dimension can be calculated by adding the lowest score in the corresponding score range of the indicator to the matching degree of the script template.

[0108] It should be noted that scores for each dimension can be limited to a range of zero to ten. The overall score can be calculated as the sum of the products of each dimension's score and its corresponding weight coefficient.

[0109] For example, for the business consulting category, the weighting coefficients for service attitude, business professionalism, communication, problem-solving, service efficiency, and risk compliance can be set to 5%, 30%, 10%, 35%, 10%, and 10%, respectively; for the standard process category, the corresponding weighting coefficients can be set to 10%, 10%, 10%, 30%, 10%, and 30%, respectively.

[0110] Optionally, the scoring process can also output non-compliant indicators, reasons for the problems, rectification suggestions, and evidence snippets. Evidence snippets are standardized text from customer service dialogue data that supports the corresponding scoring results.

[0111] Optionally, after step S102, a quality inspection report can be generated based on the quality inspection results to provide a structured summary and visual display of the results, offering a basis for customer service quality analysis, rectification and optimization, and operational management. Therefore, another embodiment of this application provides a method for generating a quality inspection report, including processes D1 to D4.

[0112] D1: Based on customer service dialogue data and quality inspection results, construct report input data according to the preset statistical cycle.

[0113] The preset statistical periods include daily report period, weekly report period and monthly report period.

[0114] Specifically, the report input data for the daily report cycle can include customer service identifier, service platform, quality inspection date, number of conversations, conversation data for each customer service representative, and corresponding comprehensive score; the report input data for the weekly report cycle can include data from seven consecutive daily report cycles; and the report input data for the monthly report cycle can include data from multiple weekly report cycles within the corresponding month.

[0115] For example, the report input data for daily, weekly, and monthly reporting cycles is shown below:

[0116] Daily input:

[0117] Customer service representative name: [Name];

[0118] Service platform: [IM / 400];

[0119] Quality inspection date: [YYYY-MM-DD];

[0120] Total conversations: [N];

[0121] Dialogue 1: Original Dialogue Score: [0-10 average];

[0122] Dialogue 2: Original Dialogue Score: [0-10 average].

[0123] Weekly report input:=

[0124] [Daily News 1];

[0125] [Daily News 2]; ...;

[0127] (Including core fields such as date, customer service name, service platform, number of people served, daily average score, achievement rate, six-dimensional analysis, and conclusion).

[0128] Monthly report input:

[0129] [Weekly Report - Week 1]

[0130] [Weekly Report for Week 2];

[0131] [Weekly Report for Week 3];

[0132] [Weekly Report for Week 4];

[0133] (Including core fields such as weekly service recipients, weekly average score, weekly achievement rate, six-dimensional qualitative summary, trends, and rectification plan).

[0134] D2: Based on the report input data, determine the service statistics and quality inspection statistics for the corresponding statistical period, and determine the trend analysis information for the corresponding statistical period based on the time series data in the report input data.

[0135] The service statistics information may include the number of services, the number of service users, and the platform distribution; the quality inspection statistics information may include the average score, the pass rate, the highest score, the lowest score, and the score distribution of each dimension; the trend analysis information may include steady improvement, high-level fluctuations, continuous decline, low-level rebound, and abnormal change points.

[0136] For example, the weekly compliance rate can be calculated by dividing the sum of the products of the compliance rate and the corresponding service quantity for each daily reporting period by the total weekly service volume, i.e., weekly compliance rate = Σ(daily compliance rate × daily service volume) ÷ total weekly service volume × 100%, rounded to the nearest integer; the monthly compliance rate can be calculated by weighting the service quantity for each weekly reporting period.

[0137] D3: Input service statistics, quality inspection statistics, and trend analysis information into the preset structured instruction template to obtain a report generation request.

[0138] It should be noted that the structured instruction template is used to limit the large language model to generate content only based on the report input data, and to limit the output fields and data format.

[0139] D4: Input the report generation request into the large language model to obtain the quality inspection report for the corresponding statistical period.

[0140] The quality inspection report includes summary information, core indicator information (e.g., a summary of 5-8 core tags), and detailed analysis information. Detailed analysis information may include a data overview, platform information, number of users served, daily average score, compliance rate, highest / lowest score, standard process implementation, and rectification actions.

[0141] Optionally, the output of the large language model can be constrained by example constraints to a preset JSON structure, and the numerical values ​​in the output results can be checked for consistency with the input data in the report. If the values ​​are inconsistent, the quality inspection report can be regenerated or the corresponding field can be corrected using the rules module.

[0142] For example, a structured instruction template is shown below:

[0143] Example constraint sample:

[0144] # Output JSON structure (strictly follow this template)

[0145] {

[0146] "abstract": "Zhang San (IM) served 5 people, with an average score of 8.4. His response was efficient and compliant, but his empathy and closed-loop management need improvement."

[0147] "core_index":["5 calls received", "Average score 8.4", "Highly efficient response", "Insufficient empathy", "Weak closed-loop feedback"],

[0148] "detail_analysis":"##Data Overview\n" Customer Service: Zhang San | Platform: IM Number of people served: 5, average daily score: 8.4, pass rate: 80%, highest / lowest: 9.0 / 7.2. Overall, the performance is at a satisfactory level, and the basic process is executed stably. ##Score Distribution | Level | Range | Number of People | Percentage | Typical Behavioral Characteristics | |---|---|---|---|---|---| | Excellent | ≥9.0 | 1 | 20% | Balanced across six dimensions, strong proactive service awareness | | Satisfactory | 8.0-8.9 | 3 | 60% | Compliant process but occasional redundant expression | | Needs Improvement | 6.0-7.9 | 1 | 20% | Insufficient empathy, positioning deviation in complex scenarios | ##Six-Dimensional Qualitative Analysis Service attitude The tone was calm but lacked empathy and proactive reassurance. Business Professional Basic queries are accurate, but location accuracy is poor in complex scenarios. Communication and Expression The logic is coherent, but the expression is redundant, and there are occasional irrelevant answers. Problem solved Standard requests can be addressed in a closed loop, while disputed solutions require further follow-up. Service efficiency The initial response met the requirements, and the response interval was reasonable. Risk Compliance No violations throughout, and standardized scripts. ##SOP Refinement Standardized Response Process: 1. Confirm the request and obtain the order number within 30 seconds; 2. Transparently inform the client that "We are relaying instructions from headquarters"; 3. Closed-loop confirmation of the result: "Effective, please check for updates." ##Rectification Actions | Priority | Improvement Direction | Specific Actions | Acceptance Standards | |---|---|---|---|---| |P0 | Enhance Empathy | Refine soothing scripts for high-emotional scenarios | Randomly check the first sentence containing emotionally resonant words the next day | |P1 | Complex Closed-Loop | Prohibit shirking responsibility for issues not attributable to this platform | Add a "continuous follow-up" commitment | ##Quality Inspection Conclusion The current service level is stable at the qualified level, with solid execution capabilities. The core risk is falling into the trap of "correct procedure but poor experience" when dealing with anxious users. Training needs to shift towards scenario simulations to strengthen empathy and adaptability.

[0149] }

[0150] It should be noted that structured instruction templates are used to ensure that the model performs qualitative inferences based solely on the input data, thus preventing data tampering and content illusion.

[0151] S103: Based on the quality inspection results, filter out the target dialogue data from the customer service dialogue data.

[0152] Among them, the target dialogue data is customer service dialogue data that meets the knowledge extraction conditions.

[0153] Specifically, knowledge extraction criteria can include dialogue categories being business consultation categories and overall scores reaching a preset scoring threshold.

[0154] For example, the preset scoring threshold can be set to 8.0 points. When the dialogue category is a standard process category or the overall score is below 8.0 points, it will not enter the candidate question-answer pair extraction process.

[0155] Optionally, knowledge extraction criteria may also include a risk compliance score reaching a compliance threshold, a problem-solving dimension indicating a closed-loop dialogue, and the absence of un-anonymized privacy information in customer service dialogue data.

[0156] S104: Extract multiple candidate question-answer pairs from the target dialogue data.

[0157] Each candidate question-answer pair includes a candidate question and a corresponding candidate answer.

[0158] It should be noted that candidate questions are used to summarize the content of users' business inquiries, and candidate answers are used to summarize the effective handling results or standardized business guidance provided by customer service.

[0159] Specifically, the target dialogue data can be input into a preset question-and-answer extraction instruction template, and the large language model can strip away greetings, emotional guidance, and redundant interactive content to extract one or two sets of candidate question-and-answer pairs.

[0160] Optionally, candidate questions can be interrogative or declarative questions that can independently express business needs, and candidate answers can be objectively expressed in the third person and converted into operational steps, rule descriptions, or standard business guidelines.

[0161] In addition, after candidate question-answer pairs are generated, privacy information identification and desensitization can be performed to remove mobile phone numbers, ID card numbers, addresses, order numbers or other user-specific information, and similar questions can be forcibly merged and deduplicated.

[0162] Optionally, each candidate question and answer pair can be associated with the source dialogue identifier, customer service identifier, quality inspection result, extraction time, and extraction confidence level to facilitate knowledge source tracing.

[0163] S105: Based on the semantic relationship between multiple candidate questions, group the multiple candidate question-answer pairs to obtain at least one question-answer group.

[0164] It should be noted that semantic association is used to characterize whether different candidate questions express the same or similar business requirements, and question-answer groups are used to aggregate candidate question-answer pairs with similar semantics.

[0165] Optionally, in another embodiment of this application, the specific implementation of step S105 includes processes E1 to E5.

[0166] E1: Vectorize multiple candidate problems to obtain the problem vectors corresponding to the multiple candidate problems.

[0167] Specifically, candidate questions and existing questions in the customer service knowledge base can be input into a pre-trained text embedding model, and high-dimensional feature encoding can be performed in a unified semantic space to generate question vectors.

[0168] For example, the pre-trained text embedding model can be the bge-m3 model, or other text embedding models that can generate semantic vectors.

[0169] E2: Determine the semantic relationship between corresponding candidate questions based on the similarity between any two question vectors.

[0170] Specifically, cosine similarity can be used as a similarity index. When the cosine similarity reaches a preset association threshold, a semantic association is determined between the corresponding candidate questions.

[0171] E3: Based on the semantic relationships between multiple candidate questions, cluster the multiple candidate questions to obtain at least one candidate question group.

[0172] For example, density-based clustering algorithms can be used for clustering. The clustering algorithm identifies high-density semantic regions based on neighborhood radius and minimum sample size, groups candidate questions in the same high-density semantic region into the same candidate question group, and marks low-density outliers as noise questions.

[0173] Optionally, the density-based clustering algorithm can be the DBSCAN algorithm, and the clustering label for noise problems can be set to -1.

[0174] E4: Select the candidate question-answer pairs corresponding to the candidate questions in each candidate question group from multiple candidate question-answer pairs.

[0175] E5: For each candidate question group, the corresponding candidate question-answer pairs are grouped into a question-answer group.

[0176] In addition, candidate question-answer pairs marked as noise issues can be grouped separately for review, or clustered again after reaching a preset number.

[0177] S106: Based on the semantic features of each candidate question in each question-answering group, determine the core representative question from each question-answering group.

[0178] Among them, the core representative question is used to represent multiple semantically similar candidate questions in the corresponding question-and-answer group.

[0179] S107: Update the customer service knowledge base based on the core representative questions and their corresponding candidate solutions.

[0180] Optionally, after step S107, a customer service knowledge base feedback optimization process can also be executed. The actual service effect after the knowledge item goes online can be used as the feedback basis for the quality inspection scoring parameters. The quality inspection scoring results can be reverse verified and dynamically corrected to improve the matching degree between the quality inspection scoring parameters and the actual service effect, thereby improving the accuracy of target dialogue screening and knowledge extraction. Therefore, in another embodiment of this application, a quality inspection scoring parameter adjustment method is provided, including processes F1 to F5.

[0181] F1: Retrieve knowledge recall feedback data corresponding to multiple knowledge items in the customer service knowledge base.

[0182] Among them, knowledge recall feedback data is used to characterize the service effect after the corresponding knowledge item is invoked, which may include recall frequency, service success rate, follow-up inquiry rate, and transfer to human agent rate.

[0183] It should be noted that the recall frequency is the number of times the knowledge item is invoked by the agent within a unit period; the service success rate can be determined based on the proportion of the number of times the user confirms that the problem is solved or no longer asks follow-up questions after the knowledge item is invoked; the follow-up question rate can be determined based on the proportion of the number of times the user asks follow-up questions continuously for a preset number of rounds (e.g., 2 times) after the knowledge item is invoked; and the transfer to human agent rate can be determined based on the proportion of the number of times the user is transferred to human agent service after the knowledge item is invoked.

[0184] F2: Based on the knowledge recall feedback data corresponding to multiple knowledge items, filter out service failure knowledge items from multiple knowledge items.

[0185] Specifically, when the service success rate is lower than the preset success rate threshold, the follow-up inquiry rate is higher than the preset follow-up inquiry rate threshold, or the rate of transferring to human agent is higher than the preset rate of transferring to human agent, the corresponding knowledge item can be identified as a service failure knowledge item.

[0186] F3: Based on the service failure knowledge entries, determine the target customer service dialogue data on which the service failure knowledge entries are based.

[0187] Specifically, based on the knowledge source identifier of the service failure knowledge entry, the corresponding source dialogue identifier is queried, and the target customer service dialogue data is determined accordingly.

[0188] F4: Retrieve the quality inspection score record corresponding to the target customer service dialogue data.

[0189] The quality inspection scoring record includes scores for each dimension, overall score, target quality inspection indicator values, and corresponding weight coefficients.

[0190] F5: Adjust the quality inspection scoring parameters based on the quality inspection scoring records.

[0191] Specifically, scores across various dimensions can be compared to identify the highest-scoring dimension in the quality inspection scoring records, and this highest-scoring dimension will be designated as the dimension to be verified. When the same dimension repeatedly becomes the highest-scoring dimension in the quality inspection scoring records corresponding to multiple service failure knowledge items, and the number of repetitions exceeds a preset threshold, it is determined that the dimension has an excessively high score.

[0192] Optionally, the threshold values ​​for dimensions with high scores can be increased by a preset step size (e.g., +0.1). When the threshold value reaches the upper limit of the corresponding score range, the threshold value is set to a saturation value, and the weight coefficient of that dimension is reduced.

[0193] In addition, to keep the sum of the weight coefficients of each dimension unchanged, the reduced weights can be allocated to other dimensions with lower current weight coefficients or those related to the cause of service failure.

[0194] For example, the adjustment step size of the indicator threshold can be set to 0.1, and the adjustment step size of the weight coefficient can be determined based on the number of quality inspection samples and the number of service failures.

[0195] It should be noted that the adjusted quality inspection scoring parameters are used for subsequent customer service dialogue data scoring processing, so that the quality inspection results can be iterated based on the actual recall effect of the customer service knowledge base.

[0196] Optionally, after parameter adjustment, offline playback can be performed using historical customer service dialogue data to compare the target dialogue filtering results and service failure knowledge item hit rate before and after adjustment, and the adjusted quality inspection scoring parameters can be released after the preset verification conditions are met.

[0197] In summary, by uniformly processing the raw customer service dialogue data of instant messaging text and voice types, customer service dialogue data with role identification is formed; quality inspection results are obtained based on dialogue categories and target quality inspection indicators, and candidate question-answer pairs are extracted from the target dialogue data that meet the screening conditions; synonymous questions are grouped and knowledge entries are updated through semantic association, question-answer groups, and core representative questions; and then the quality inspection scoring records are traced back and the quality inspection scoring parameters are adjusted using knowledge retrieval feedback data, thereby forming a feedback loop of customer service dialogue quality inspection, knowledge extraction, customer service knowledge base update, and quality inspection scoring parameter optimization.

[0198] like Figure 2 The flowchart shown illustrates another knowledge base optimization method based on customer service dialogue quality inspection, which includes the following steps:

[0199] S201: Obtain customer service dialogue data and customer service knowledge base.

[0200] It should be noted that the specific implementation of step S201 can be referred to step S101 accordingly, and will not be repeated here.

[0201] S202: Perform quality inspection on customer service dialogue data and obtain the quality inspection results.

[0202] It should be noted that the specific implementation of step S202 can be referred to step S102 accordingly, and will not be repeated here.

[0203] S203: Based on the quality inspection results, select target dialogue data from customer service dialogue data.

[0204] It should be noted that the specific implementation of step S203 can be found in step S103, and will not be repeated here.

[0205] S204: Extract multiple candidate question-answer pairs from the target dialogue data.

[0206] Each candidate question-answer pair includes a candidate question and a corresponding candidate answer.

[0207] It should be noted that the specific implementation of step S204 can be found in step S104, and will not be repeated here.

[0208] S205: Based on the semantic relationship between multiple candidate questions, group the multiple candidate question-answer pairs to obtain at least one question-answer group.

[0209] It should be noted that the specific implementation of step S205 can be referred to step S105 accordingly, and will not be repeated here.

[0210] S206: Vectorize each candidate question in each question-answering group to obtain the question vector corresponding to each candidate question.

[0211] It should be noted that when the question vector corresponding to the candidate question has been obtained in step S205, the corresponding question vector can be reused directly.

[0212] S207: Determine the center vector of the corresponding question and answer group based on the average value of the question vectors in each question and answer group.

[0213] Specifically, the central vector is obtained by taking the arithmetic mean of the values ​​of each question vector in each vector dimension.

[0214] S208: Calculate the feature distance between the question vector and the center vector for each candidate question.

[0215] For example, the feature distance can be either Euclidean distance or cosine distance.

[0216] S209: Compare the feature distances corresponding to each candidate problem.

[0217] It should be noted that the candidate question with the minimum feature distance is the closest to the center position in the corresponding semantic cluster and can represent the main semantics of the question-answer group.

[0218] Optionally, when multiple candidate questions have the same feature distance, the core representative question can be determined based on the quality inspection results, extraction confidence, text integrity, and the number of source dialogues.

[0219] S210: The candidate question with the smallest feature distance is identified as the core representative question of the corresponding question-answering group.

[0220] S211: Update the customer service knowledge base based on the core representative questions and their corresponding candidate solutions.

[0221] It should be noted that the specific implementation of step S211 can be found in step S107, and will not be repeated here.

[0222] In summary, by vectorizing the candidate questions in each question-and-answer group and determining the center vector of the group based on the average of the question vectors corresponding to each candidate question, and then determining the core representative questions based on the feature distance between each candidate question and the center vector, the determined core representative questions can more accurately represent the overall semantics of the corresponding question-and-answer group. Compared with methods such as random selection, selection based on frequency of occurrence, or selection based on time order, this embodiment can reduce the impact of individual abnormal expressions, redundant expressions, or marginal semantics on the knowledge accumulation results, improve the representativeness and stability of the core representative questions, further improve the quality of knowledge item generation, reduce duplicate knowledge and semantic offset problems, and improve the accuracy of customer service knowledge base updates and knowledge reuse efficiency.

[0223] like Figure 3 The flowchart shown illustrates another knowledge base optimization method based on customer service dialogue quality inspection, which includes the following steps:

[0224] S301: Obtain customer service dialogue data and customer service knowledge base.

[0225] It should be noted that the specific implementation of step S301 can be found in step S101, and will not be repeated here.

[0226] S302: Perform quality inspection on customer service dialogue data and obtain the quality inspection results.

[0227] It should be noted that the specific implementation of step S302 can be referred to step S102 accordingly, and will not be repeated here.

[0228] S303: Based on the quality inspection results, filter out the target dialogue data from the customer service dialogue data.

[0229] It should be noted that the specific implementation of step S303 can be found in step S103, and will not be repeated here.

[0230] S304: Extract multiple candidate question-answer pairs from the target dialogue data.

[0231] Each candidate question-answer pair includes a candidate question and a corresponding candidate answer;

[0232] It should be noted that the specific implementation of step S304 can be referred to step S104 accordingly, and will not be repeated here.

[0233] S305: Based on the semantic relationship between multiple candidate questions, group multiple candidate question-answer pairs to obtain at least one question-answer group.

[0234] It should be noted that the specific implementation of step S305 can be referred to step S105 accordingly, and will not be repeated here.

[0235] S306: Based on the semantic features of each candidate question in each question-answering group, determine the core representative question from each question-answering group.

[0236] It should be noted that the specific implementation of step S306 can be found in step S106, and will not be repeated here.

[0237] S307: Select candidate answers for the core representative question from multiple candidate question-answer pairs.

[0238] S308: Perform semantic consistency verification on core representative questions and candidate solutions.

[0239] Specifically, large language models or natural language inference models can be used to determine whether candidate solutions can answer the core representative question and to check whether there are any business facts in the candidate solutions that are inconsistent with the source dialogue.

[0240] S309: If the core representative question and the candidate solution have the same semantics, then generate the target knowledge entry based on the core representative question and the candidate solution.

[0241] The target knowledge items may include standard questions, standard answers, knowledge categories, scope of application, knowledge tags, and update time.

[0242] S310: Add a knowledge source identifier to the target knowledge entry.

[0243] The knowledge source identifier is used to indicate that the target knowledge item was extracted from customer service dialogue data. The knowledge source identifier can also be associated with the source dialogue identifier, quality inspection results, and the question-and-answer group identifier corresponding to the core representative question.

[0244] S311: Write the target knowledge entry with the added knowledge source identifier into the customer service knowledge base.

[0245] Optionally, target knowledge entries can be written into the pending review area first, and then switched to the recallable state after passing manual review or automatic consistency verification.

[0246] In addition, when there are already knowledge entries in the customer service knowledge base that are semantically similar to the core representative question, the candidate answer can be compared with the existing answer for consistency, and the original knowledge entry can be merged, supplemented, updated or kept based on the comparison results.

[0247] In summary, by performing semantic consistency checks on core representative questions and candidate solutions before knowledge is added to the knowledge base, target knowledge entries are generated only when they are semantically consistent. These target knowledge entries are then written into the customer service knowledge base after a knowledge source identifier is added. This ensures that the generated knowledge entries not only have high semantic accuracy but also completely record the knowledge source information, achieving end-to-end traceability management of knowledge entries. When semantically similar knowledge entries exist in the customer service knowledge base, consistency comparisons can be used to perform knowledge merging, supplementation, or version updates. This prevents erroneous, conflicting, and duplicate knowledge from entering the customer service knowledge base, improving the consistency, reliability, and maintainability of the knowledge base content. It also provides reliable data support for subsequent knowledge review, iteration, and retrieval effect analysis, further enhancing the application quality and continuous optimization capabilities of the intelligent customer service knowledge base.

[0248] like Figure 4 The diagram shown is an architectural schematic of a knowledge base optimization device based on customer service dialogue quality inspection provided in this application embodiment. The optimization device includes: an acquisition unit 100, a quality inspection unit 200, a filtering unit 300, an extraction unit 400, a grouping unit 500, a determination unit 600, and an update unit 700.

[0249] Unit 100 is used to acquire customer service dialogue data and customer service knowledge base.

[0250] Acquisition unit 100 includes:

[0251] The data acquisition subunit is used to acquire raw customer service dialogue data.

[0252] The first determining subunit is used to determine the data type corresponding to the original customer service dialogue data. The data types include instant messaging text type and voice type.

[0253] The conversion subunit is used to convert the original customer service dialogue data into speech if the data type is speech, so as to obtain the original transcribed text.

[0254] The first extraction subunit is used to extract the original dialogue text from the original customer service dialogue data if the data type is instant messaging text.

[0255] The second extraction subunit is used to extract multiple statements from the original transcribed text or the original dialogue text.

[0256] The second determining subunit is used to determine the speaking role corresponding to each statement based on the dialogue context corresponding to each statement.

[0257] Add a sub-unit to add the role identifier corresponding to the speaking role to the corresponding speaking content to obtain the role label text.

[0258] The processing subunit is used to standardize the character annotation text to obtain standardized text.

[0259] The combination subunit is used to combine the standardized texts according to the time sequence corresponding to each standardized text to obtain customer service dialogue data.

[0260] Quality inspection unit 200 is used to perform quality inspection on customer service dialogue data and obtain quality inspection results.

[0261] Quality inspection unit 200 includes:

[0262] The third sub-unit is used to determine the dialogue category corresponding to the customer service dialogue data; the dialogue categories include business consultation category and standard process category.

[0263] The third determining subunit is specifically used for: performing intent recognition on customer service dialogue data to obtain dialogue intent; extracting features from the dialogue intent to obtain dialogue intent features; obtaining rule intent features corresponding to multiple rule intents in the rule base; calculating the similarity between the dialogue intent features and the rule intent features to obtain intent matching results; and determining the dialogue category corresponding to the customer service dialogue data based on the intent matching results.

[0264] The fourth determination subunit is used to determine the target quality inspection indicator corresponding to the dialogue category from multiple quality inspection indicators based on the dialogue category.

[0265] The quality inspection subunit is used to score customer service dialogue data based on target quality inspection indicators to obtain quality inspection results.

[0266] The filtering unit 300 is used to filter target dialogue data from customer service dialogue data based on the quality inspection results.

[0267] Extraction unit 400 is used to extract multiple candidate question-answer pairs from the target dialogue data; each candidate question-answer pair includes a candidate question and a corresponding candidate answer.

[0268] Grouping unit 500 is used to group multiple candidate question-answer pairs based on the semantic relationship between multiple candidate questions to obtain at least one question-answer group.

[0269] The grouping unit 500 is specifically used for: vectorizing multiple candidate questions to obtain question vectors corresponding to multiple candidate questions; determining the semantic relationship between corresponding candidate questions based on the similarity between any two question vectors; clustering multiple candidate questions based on the semantic relationship between multiple candidate questions to obtain at least one candidate question group; selecting candidate question-answer pairs corresponding to candidate questions in each candidate question group from multiple candidate question-answer pairs; and forming a question-answer group for each candidate question group by combining the corresponding candidate question-answer pairs.

[0270] The determination unit 600 is used to determine the core representative question from each question-answering group based on the semantic features of each candidate question in each question-answering group.

[0271] The determining unit 600 is specifically used for: vectorizing each candidate question in each question-answer group to obtain the question vector corresponding to each candidate question; determining the center vector of the corresponding question-answer group based on the average value of each question vector in each question-answer group; calculating the feature distance between the question vector corresponding to each candidate question and the center vector; comparing the feature distances corresponding to each candidate question; and determining the candidate question with the smallest feature distance as the core representative question of the corresponding question-answer group.

[0272] Update Unit 700 is used to update the customer service knowledge base based on the core representative questions and their corresponding candidate solutions.

[0273] The update unit 700 is specifically used for: selecting candidate answers corresponding to core representative questions from multiple candidate question-answer pairs; performing semantic consistency verification on core representative questions and candidate answers; if the core representative questions and candidate answers are semantically consistent, generating target knowledge entries based on the core representative questions and candidate answers; adding knowledge source identifiers to target knowledge entries; the knowledge source identifiers are used to indicate that the target knowledge entries are extracted from customer service dialogue data; and writing the target knowledge entries with added knowledge source identifiers into the customer service knowledge base.

[0274] In summary, by outputting structured quality inspection results through quality inspection processing, and using these results to filter target dialogue data from the original dialogue, the barrier between independent quality inspection and knowledge updating is broken down, making the quality inspection process a direct trigger for knowledge accumulation. Subsequently, candidate question-answer pairs are extracted from the target dialogue, and grouped using the semantic relationships between these candidate questions to form question-answer groups. Then, core representative questions are extracted based on the semantic features within each group, ensuring the reduction of redundancy and semantic coverage of knowledge entries. Finally, the customer service knowledge base is dynamically updated with the core representative questions and their corresponding candidate answers, achieving quality inspection-driven knowledge increment injection. Simultaneously, after the customer service knowledge base is updated, the coverage deviation of the initial quality inspection rules can be verified, providing data support for subsequent quality inspection standard optimization. This forms a closed-loop feedback mechanism from quality inspection to knowledge extraction, then to customer service knowledge base updates, and finally to quality inspection rule optimization. This fundamentally overcomes the shortcomings of existing technologies where quality inspection and customer service knowledge bases operate independently and cannot evolve collaboratively, continuously improving the self-optimization capability of the customer service system and the timeliness and accuracy of the customer service knowledge base.

[0275] Combination Figure 4 The optimization device, as shown, also includes:

[0276] The construction unit is used to build report input data based on customer service dialogue data and quality inspection results, according to a preset statistical cycle. The preset statistical cycle includes daily report cycle, weekly report cycle and monthly report cycle.

[0277] The information determination unit is used to determine the service statistics and quality inspection statistics for the corresponding statistical period based on the report input data, and to determine the trend analysis information for the corresponding statistical period based on the time series data in the report input data.

[0278] The input unit is used to input service statistics, quality inspection statistics, and trend analysis information into a preset structured instruction template to obtain a report generation request.

[0279] The report input unit is used to input the report generation request into the large language model to obtain the quality inspection report for the corresponding statistical period; the quality inspection report includes summary information, core indicator information and detailed analysis information.

[0280] Combination Figure 4 The optimization device, as shown, also includes:

[0281] The feedback data acquisition unit is used to acquire knowledge retrieval feedback data corresponding to multiple knowledge items in the customer service knowledge base. The knowledge retrieval feedback data is used to characterize the service effect after the corresponding knowledge item is invoked.

[0282] The item unit is used to filter out service failure knowledge items from multiple knowledge items based on the knowledge recall feedback data corresponding to multiple knowledge items.

[0283] The data determination unit is used to determine the target customer service dialogue data on which the service failure knowledge entries are based, based on the service failure knowledge entries.

[0284] The record acquisition unit is used to acquire the quality inspection score records corresponding to the target customer service dialogue data.

[0285] The adjustment unit is used to adjust the quality inspection scoring parameters based on the quality inspection scoring records.

[0286] Another embodiment of this application provides an electronic device, such as... Figure 5 As shown, it includes: memory 501, processor 502 and bus 503.

[0287] The memory 501 is used to store the program.

[0288] The processor 502 is used to execute a program, which, when executed, is specifically used to implement a knowledge base optimization method based on customer service dialogue quality inspection as provided in any of the above embodiments.

[0289] The electronic devices mentioned in this article can be servers, PCs, tablets, mobile phones, ECUs (Electronic Control Units), VCUs (Vehicle Control Units), MCUs (Micro Controller Units), HCUs (Hybrid Control Units), etc.

[0290] Another embodiment of this application provides a computer storage medium for storing a computer program, which, when executed, implements a knowledge base optimization method based on customer service dialogue quality inspection as provided in any of the above embodiments.

[0291] Computer storage media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0292] This application also provides a computer program product comprising executable computer instructions, which, when executed by a processor of a computer device, are used to implement a knowledge base optimization method based on customer service dialogue quality inspection as provided in any of the above embodiments.

[0293] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0294] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing a knowledge base based on quality inspection of customer service dialogues, characterized in that, include: Acquire customer service conversation data and customer service knowledge base; The customer service dialogue data is subjected to quality inspection to obtain the quality inspection results; Based on the quality inspection results, target dialogue data is selected from the customer service dialogue data; Multiple candidate question-answer pairs are extracted from the target dialogue data; each candidate question-answer pair includes a candidate question and a corresponding candidate answer. Based on the semantic relationships between the multiple candidate questions, the multiple candidate question-answer pairs are grouped to obtain at least one question-answer group; Based on the semantic features of each candidate question in each question-answer group, a core representative question is determined from each question-answer group; The customer service knowledge base is updated based on the core representative questions and their corresponding candidate solutions.

2. The method according to claim 1, characterized in that, The process of obtaining customer service conversation data includes: Obtain raw customer service conversation data; Determine the data type corresponding to the original customer service dialogue data, including instant messaging text type and voice type; If the data type is the voice type, then the original customer service dialogue data is converted into voice to obtain the original transcribed text; If the data type is the instant messaging text type, then the original dialogue text is extracted from the original customer service dialogue data; Extract multiple statements from the original transcribed text or the original dialogue text; Based on the dialogue context corresponding to each of the aforementioned statements, determine the speaking role corresponding to each of the aforementioned statements. Add the role identifier corresponding to the speaking role to the corresponding speaking content to obtain the role annotation text; The character annotation text is standardized to obtain standardized text; According to the time sequence corresponding to each of the standardized texts, the standardized texts are combined to obtain customer service dialogue data.

3. The method according to claim 1, characterized in that, The quality inspection process for the customer service dialogue data, to obtain the quality inspection results, includes: Determine the dialogue category corresponding to the customer service dialogue data; the dialogue category includes business consultation category and standard process category. Based on the dialogue category, a target quality inspection indicator corresponding to the dialogue category is determined from multiple quality inspection indicators; Based on the target quality inspection indicators, the customer service dialogue data is scored to obtain the quality inspection results.

4. The method according to claim 3, characterized in that, Determining the dialogue category corresponding to the customer service dialogue data includes: The customer service dialogue data is subjected to intent recognition to obtain the dialogue intent; The dialogue intent is feature extracted to obtain dialogue intent features; Obtain the rule intent features corresponding to multiple rule intents in the rule base; The similarity between the dialogue intent features and the rule intent features is calculated to obtain the intent matching result; The dialogue category corresponding to the customer service dialogue data is determined based on the intent matching result.

5. The method according to claim 1, characterized in that, Based on the semantic relationships between the candidate questions, the candidate question-answer pairs are grouped to obtain at least one question-answer group, including: The candidate questions are vectorized to obtain question vectors corresponding to the candidate questions. Based on the similarity between any two question vectors, the semantic association between the corresponding candidate questions is determined; Based on the semantic relationships between the multiple candidate questions, the multiple candidate questions are clustered to obtain at least one candidate question group; From the multiple candidate question-answer pairs, select the candidate question-answer pairs corresponding to the candidate questions in each candidate question group; For each candidate question group, the corresponding candidate question-answer pairs are grouped into a question-answer group.

6. The method according to claim 1, characterized in that, The step of determining the core representative question from each question-answering group based on the semantic features of each candidate question in each question-answering group includes: Each candidate question in each question-answer group is vectorized to obtain a question vector corresponding to each candidate question. Based on the average value of each question vector in each question-answer group, determine the center vector corresponding to the question-answer group; Calculate the feature distance between the question vector corresponding to each candidate question and the center vector; The feature distances corresponding to each candidate problem are compared; The candidate question with the smallest feature distance is determined as the core representative question of the corresponding question-answer group.

7. The method according to claim 1, characterized in that, The step of updating the customer service knowledge base based on the core representative question and the corresponding candidate answers includes: From the multiple candidate question-answer pairs, candidate answers corresponding to the core representative question are selected; Perform semantic consistency verification on the core representative question and the candidate solutions; If the core representative question and the candidate solution have the same semantics, then a target knowledge entry is generated based on the core representative question and the candidate solution. Add a knowledge source identifier to the target knowledge entry; the knowledge source identifier is used to indicate that the target knowledge entry was extracted from customer service dialogue data. The target knowledge entries, after being marked with a knowledge source identifier, are written into the customer service knowledge base.

8. The method according to claim 1, characterized in that, After performing quality inspection on the customer service dialogue data and obtaining the quality inspection results, the process further includes: Based on the customer service dialogue data and the quality inspection results, report input data is constructed according to a preset statistical cycle, wherein the preset statistical cycle includes a daily report cycle, a weekly report cycle, and a monthly report cycle; Based on the report input data, service statistics and quality inspection statistics for the corresponding statistical period are determined, and trend analysis information for the corresponding statistical period is determined based on the time series data in the report input data. Input the service statistics, quality inspection statistics, and trend analysis information into a preset structured instruction template to obtain a report generation request; The report generation request is input into the large language model to obtain the quality inspection report for the corresponding statistical period; the quality inspection report includes summary information, core indicator information and detailed analysis information.

9. The method according to claim 1, characterized in that, Also includes: Obtain knowledge recall feedback data corresponding to multiple knowledge items in the customer service knowledge base. The knowledge recall feedback data is used to characterize the service effect after the corresponding knowledge item is invoked. Based on the knowledge recall feedback data corresponding to multiple knowledge items, service failure knowledge items are selected from multiple knowledge items; Based on the service failure knowledge entries, determine the target customer service dialogue data on which the service failure knowledge entries are generated; Obtain the quality inspection score record corresponding to the target customer service dialogue data; Based on the quality inspection scoring records, the quality inspection scoring parameters are adjusted.

10. A knowledge base optimization device based on customer service dialogue quality inspection, characterized in that, include: The acquisition unit is used to acquire customer service dialogue data and customer service knowledge base; The quality inspection unit is used to perform quality inspection processing on the customer service dialogue data and obtain the quality inspection results. A filtering unit is used to filter target dialogue data from the customer service dialogue data based on the quality inspection results; An extraction unit is used to extract multiple candidate question-answer pairs from the target dialogue data; each candidate question-answer pair includes a candidate question and a corresponding candidate answer. A grouping unit is used to group multiple candidate question-answer pairs based on the semantic association between multiple candidate questions to obtain at least one question-answer group. A determining unit is configured to determine a core representative question from each of the question-answering groups based on the semantic features of each candidate question in each question-answering group; The update unit is used to update the customer service knowledge base based on the core representative question and the candidate answers corresponding to the core representative question.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed by a processor, performs the knowledge base optimization method based on customer service dialogue quality inspection as described in any one of claims 1-9.

12. An electronic device, characterized in that, include: Processor, memory, and bus; The processor and the memory are connected via the bus; The memory is used to store the program, and the processor is used to run the program, wherein the program is executed by the processor to perform the knowledge base optimization method based on customer service dialogue quality inspection as described in any one of claims 1-9.

13. A computer program product, characterized in that, It includes executable computer instructions, which, when executed by a processor of a computer device, are used to implement the knowledge base optimization method based on customer service dialogue quality inspection as described in any one of claims 1-9.

Images