Complaint work order processing method and device, electronic equipment and storage medium

Abstract

The invention provides a complaint work order processing method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a to-be-processed complaint work order, and extracting the key features of the complaint work order; inputting the key features of the complaint work order into a pre-trained first large language model to obtain a complaint work order processing department category output by the first large language model; inputting the key features of the complaint work order into a pre-trained second large language model to obtain a complaint work order processing strategy output by the second large language model; wherein the first large language model is obtained by training based on sample key features of the sample complaint work order and processing department category labels of the sample complaint work order; the second large language model is obtained by training based on the sample key features of the sample complaint work order and the processing strategy label of the sample complaint work order. According to the invention, the processing strategy of the complaint work order can be automatically generated, and the complaint processing efficiency and quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, electronic device and storage medium for processing complaint work orders. Background Art

[0002] With the rapid development of the service industry, customers' expectations for service quality are increasing, and the number of customer complaints is on the rise. Handling customer complaints is a crucial element in ensuring service quality and safeguarding consumer rights. Currently, existing technologies primarily focus on the generation and flow of complaint tickets, making it difficult to automate the generation of complaint handling strategies. Summary of the Invention

[0003] The present invention provides a complaint worksheet processing method, device, electronic device and storage medium to solve the problem that the existing technology mainly focuses on studying the generation and circulation of complaint worksheets and is difficult to realize the automatic generation of complaint handling strategies.

[0004] The present invention provides a method for processing a complaint work order, comprising: Obtain a pending complaint ticket and extract key features of the complaint ticket; Inputting the key features of the complaint worksheet into a pre-trained first language model to obtain the processing department category of the complaint worksheet output by the first language model; Inputting the key features of the complaint ticket into a pre-trained second language model to obtain a processing strategy for the complaint ticket output by the second language model; Among them, the first language model is trained based on the sample key features of the sample complaint work orders and the processing department category labels of the sample complaint work orders; the second language model is trained based on the sample key features of the sample complaint work orders and the processing strategy labels of the sample complaint work orders.

[0005] In some embodiments, the second largest language model includes an extraction layer, a matching layer, and a decision layer; Correspondingly, inputting the key features of the complaint ticket into the pre-trained second language model to obtain the processing strategy for the complaint ticket output by the second language model includes: Inputting the key features of the complaint worksheet into the extraction layer to obtain the attribute features of each entity of the complaint worksheet output by the extraction layer; Inputting the attribute characteristics of each entity in the complaint ticket into the matching layer, the matching layer matches the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in the pre-built regulatory map, and determining the guiding measures for the complaint ticket based on the matching results; The guiding measures of the complaint worksheet are input into the decision-making layer, and a handling strategy for the complaint worksheet output by the decision-making layer is obtained. The handling strategy includes an apology, compensation measures and a prevention strategy.

[0006] In some embodiments, the process of constructing the regulatory map includes: Building a regulatory knowledge base, wherein the regulatory knowledge base stores multiple regulatory documents corresponding to complaint work orders; Extracting the entity-relationship-entity in each regulatory document, and constructing an initial regulatory graph based on the entity-relationship-entity in each regulatory document; Clustering and screening each entity in the initial regulatory map to obtain an optimized regulatory map, and filling in the attributes of each entity in the optimized regulatory map to obtain the regulatory map; Determining, based on the multiple regulatory documents, guiding measures corresponding to the attributes of each entity in the regulatory map; Feature extraction is performed on the attributes of each entity in the regulatory law map to obtain attribute features of each entity in the regulatory law map.

[0007] In some embodiments, matching the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in a pre-built regulatory map, and determining guidance measures for the complaint ticket based on the matching results, includes: The attribute characteristics of each entity of the complaint worksheet are precisely matched with the attribute characteristics of each entity of the regulatory law map to determine whether the match is successful. If so, the guidance measures for the complaint worksheet are determined based on the precise matching results. Otherwise, the attribute characteristics of each entity of the complaint worksheet are fuzzy matched with the attribute characteristics of each entity of the regulatory law map to determine whether the match is successful. If so, the guidance measures for the complaint worksheet are determined based on the fuzzy matching results. Otherwise, the recommended guidance measures for the complaint worksheet are determined.

[0008] In some embodiments, the method further comprises: Obtaining feedback data on the handling strategy of the complaint work order; Inputting feedback data of the handling strategy of the complaint ticket into the evaluation layer of the pre-trained third language model, and obtaining a quality evaluation result of the handling strategy of the complaint ticket output by the evaluation layer; Optimizing the parameters of the second language model based on the quality evaluation results of the handling strategy of the complaint ticket; Among them, the evaluation layer of the third language model is trained based on historical feedback data of historical processing strategies of historical complaint work orders.

[0009] In some embodiments, the method further comprises: Inputting the handling strategy of the complaint ticket into the decision layer of the pre-trained third language model, the decision layer determines whether the handling strategy meets the regulatory exclusion scenario, and obtains the decision result of the handling strategy of the complaint ticket; If it is determined that the handling strategy of the complaint ticket meets the regulatory elimination scenario, eliminate the handling strategy of the complaint ticket; Optimizing the parameters of the second language model according to the determination result of the handling strategy of the complaint ticket; Among them, the decision layer of the third language model is trained based on the sample processing strategy of the sample complaint work order and the decision result label of the sample processing strategy.

[0010] In some embodiments, the training process of the second language model includes: Obtaining a sample complaint work order, and extracting sample key features of the sample complaint work order; Determine a handling strategy label for the sample complaint ticket; Inputting the sample key features of the sample complaint worksheet into the initial second largest language model to obtain a processing strategy for the sample complaint worksheet output by the initial second largest language model; Based on the processing strategy of the sample complaint work order and the processing strategy label of the sample complaint work order, the loss function value is calculated, and the parameters of the initial second largest language model are iteratively optimized to obtain the second largest language model.

[0011] In some embodiments, after obtaining the second largest language model, the method further includes: Obtaining feedback data on the handling strategy of the sample complaint work order; Inputting the feedback data of the handling strategy of the sample complaint work order into the evaluation layer of the pre-trained third language model, and obtaining the quality evaluation result of the handling strategy of the sample complaint work order output by the evaluation layer; Inputting the handling strategy of the sample complaint work order into the decision layer of the pre-trained third language model, the decision layer determines whether the handling strategy of the sample complaint work order meets the regulatory exclusion scenario, and obtains the decision result of the handling strategy of the sample complaint work order; Optimize the parameters of the second language model based on the quality evaluation results of the processing strategy of the sample complaint worksheet and the determination results of the processing strategy of the sample complaint worksheet.

[0012] The present invention also provides a complaint work order processing device, comprising: A feature extraction unit is used to obtain a complaint ticket to be processed and extract key features of the complaint ticket; a classification unit, configured to input key features of the complaint ticket into a pre-trained first language model, and obtain a processing department category of the complaint ticket output by the first language model; A decision-making unit, configured to input key features of the complaint ticket into a pre-trained second language model to obtain a processing strategy for the complaint ticket output by the second language model; Among them, the first language model is trained based on the sample key features of the sample complaint work orders and the processing department category labels of the sample complaint work orders, and the second language model is trained based on the sample key features of the sample complaint work orders and the processing strategy labels of the sample complaint work orders.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for processing a complaint work order as described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any of the complaint ticket processing methods described above.

[0015] The complaint ticket processing method, device, electronic device and storage medium provided by the present invention obtain the complaint ticket to be processed, extract the key features of the complaint ticket, input the key features of the complaint ticket into a pre-trained first language model, obtain the processing department category of the complaint ticket output by the first language model, input the key features of the complaint ticket into a pre-trained second language model, and obtain the processing strategy of the complaint ticket output by the second language model. This can not only automatically classify the processing departments of the complaint ticket, but also automatically generate the processing strategy of the complaint ticket, thereby further improving the efficiency and quality of complaint processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 It is a flowchart of the complaint ticket processing method provided by an embodiment of the present invention.

[0018] Figure 2This is a flowchart of inputting key features of a complaint ticket into a pre-trained second language model provided by an embodiment of the present invention.

[0019] Figure 3 It is a flowchart of the training process of the second language model provided by an embodiment of the present invention.

[0020] Figure 4 It is a structural diagram of the complaint work order processing device provided by an embodiment of the present invention.

[0021] Figure 5 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0023] The terms "first," "second," and the like in the present invention are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention can be implemented in an order other than that illustrated or described herein. Furthermore, the objects distinguished by "first" and "second" are generally of the same type, and do not limit the number of objects. For example, the first object may be one or more.

[0024] As an important part of the service industry, the quality of financial services is directly related to customer satisfaction and the healthy development of the industry. Traditional manual processing methods have the following problems: 1) Inefficient processing: Manual complaint processing is slow and cannot quickly respond to a large number of complaints; 2) Unstable handling quality: Due to differences in customer service staff’s experience and capabilities, the quality of complaint handling decisions varies; 3) Difficulty in ensuring regulatory compliance: When manually handling complaints, it is difficult to ensure that every complaint complies with regulatory requirements; 4) Lack of effective post-evaluation mechanism: There is a lack of effective mechanism to evaluate the compliance and rationality of complaints that have been handled.

[0025] In the financial industry's customer complaint handling process, existing technologies primarily focus on the generation and circulation of work tickets, specifically the initial organization of complaint incidents and their assignment to the appropriate handling departments. However, these technical solutions primarily address the issues of inefficient complaint ticket processing and inconsistent work ticket quality, while lacking effective technical support for the subsequent handling of complaint tickets, particularly the generation of complaint handling strategies and the evaluation of their rationality. Specifically, the shortcomings of existing technologies include the following two main points: 1) Lack of complaint handling decision generation: Existing technologies do not address how to generate specific complaint handling strategies based on the content of complaint tickets. This step is a critical stage in the financial industry's complaint handling process, directly related to the protection of consumer rights and the reputation management of financial institutions. 2) Inadequacies in regulatory compliance assessments: Regulatory bodies' complaint handling regulations require multiple verifications of complaint content, and complaint handlers must be assisted in determining which exclusion scenarios the customer complaint falls under and providing exclusion application recommendations. Existing technologies fail to provide the technical support required for this process.

[0026] To this end, an embodiment of the present invention provides a method for processing complaint tickets. By extracting the key features of the complaint ticket and using a pre-trained first language model, the processing departments of the complaint ticket are automatically classified, and the pre-trained second language model is used to automatically generate a processing strategy for the complaint ticket, thereby further improving the efficiency and quality of complaint handling.

[0027] Figure 1 This is a flow chart of the method for handling complaint tickets provided by an embodiment of the present invention. Figure 1 As shown, a method for handling a complaint work order is provided, comprising the following steps: step 110, step 120, and step 130. The steps of the method flow are merely a possible implementation of the present invention.

[0028] Step 110: Obtain the complaint ticket to be processed and extract key features of the complaint ticket.

[0029] Among them, the complaint ticket is a document or electronic record used to record and manage customer complaints. The complaint ticket describes in detail the customer's dissatisfaction, the nature of the problem, and the desired solution.

[0030] Optionally, the complaint ticket is a complaint ticket from a service industry such as a financial institution or a bank.

[0031] Optionally, the complaint ticket can be obtained through various channels, such as receiving text uploaded by the user or receiving the text obtained by performing voice recognition on the user's voice data.

[0032] For example, a customer named Mr. Li called the bank's customer service hotline to file a complaint and excitedly recounted the problems he encountered when handling business at the bank. Through voice recognition technology, Mr. Li's voice content was accurately converted into complaint ticket text.

[0033] Optionally, complaint ticket data can be collected through terminal devices, which can be mobile phones, personal computers, tablet computers and other devices.

[0034] Among them, the key features include at least customer identity, complaint reasons, demands and other characteristics; customer identity includes customer ID, customer name, customer level, contact information and other information.

[0035] For example, the customer's identity is "Mr. Li", the reason for the complaint is "the credit card was not received within the promised time", and the request is "I hope to receive the credit card as soon as possible" or "get a reasonable explanation and compensation."

[0036] Optionally, the complaint ticket is preprocessed to obtain a preprocessed complaint ticket, and feature extraction is performed on the preprocessed complaint ticket to obtain key features.

[0037] In some embodiments, key features of the complaint ticket are extracted, including: Perform pre-processing on the complaint ticket, including word segmentation, part-of-speech tagging, and entity recognition, to obtain the pre-processed complaint ticket; Input the pre-processed complaint ticket into the pre-trained fourth language model to obtain the key features of the complaint ticket output by the fourth language model; Among them, the fourth language model is trained based on the preprocessed sample complaint tickets and the key feature labels of the sample complaint tickets.

[0038] Optionally, the fourth language model may be a Bidirectional Encoder Representations from Transformer (BERT) model.

[0039] For example, after word segmentation of Mr. Li's complaint ticket text, we get "I applied for a credit card at your bank last week. We were told I would receive the card within a week, but I still haven't received it! I'm still waiting to use this card!" Part-of-speech tagging can determine the part of speech of each word. For example, "I" is a pronoun and "last week" is a time noun. Entity recognition can identify key entities such as bank and credit card.

[0040] Optionally, the training process of the fourth language model includes: Obtain sample complaint tickets and determine key feature tags of the sample complaint tickets; Perform preprocessing on the sample complaint ticket, including word segmentation, part-of-speech tagging, and entity recognition, to obtain the preprocessed sample complaint ticket; Input the preprocessed sample complaint worksheet into the initial fourth language model to obtain the key features of the sample complaint worksheet output by the initial fourth language model; Based on the key features of the sample complaint tickets and the key feature labels of the sample complaint tickets, the parameters of the initial fourth language model are iteratively optimized to obtain the fourth language model.

[0041] Step 120: Input the key features of the complaint ticket into the pre-trained first language model to obtain the processing department category of the complaint ticket output by the first language model.

[0042] Among them, the first language model is trained based on the sample key features of the sample complaint tickets and the category labels of the processing departments of the sample complaint tickets.

[0043] Optionally, the key features of the complaint ticket are analyzed based on the first language model to determine the entities, business types and business demands involved in the complaint ticket. The complaint ticket is classified according to the entities, business types and business demands involved in the complaint incident, and the department category for handling the complaint ticket is determined.

[0044] For example, the key feature of a complaint ticket is "the credit card was not received within a week, and the customer requested that the credit card be sent as soon as possible." The entity-business-business demand involved is determined to be "credit card-card processing-quick card processing," and the department responsible for handling the complaint ticket can be determined to be the "credit card business processing department."

[0045] It can be understood that by inputting the key features of the complaint ticket into the pre-trained first language model and obtaining the processing department category of the complaint ticket output by the first language model, the time and cost of manual classification can be greatly reduced, thereby speeding up the entire complaint handling process, helping to ensure that the complaint ticket is correctly assigned to the relevant department, reducing delays or omissions caused by incorrect classification, and improving customer experience.

[0046] Optionally, the training process of the first language model includes: Obtain sample complaint tickets and extract key features of the sample complaint tickets; Determine the department category label for the sample complaint ticket; Input the sample key features of the sample complaint work order into the initial first language model to obtain the processing department category of the sample complaint work order output by the initial first language model; Based on the processing department category of the sample complaint work order and the processing department category label of the sample complaint work order, the parameters of the initial first language model are iteratively optimized to obtain the first language model.

[0047] It can be understood that training the initial first language model based on the sample key features of the sample complaint work orders and the processing department category labels of the sample complaint work orders to obtain the first language model can improve the classification accuracy of the first language model.

[0048] In some embodiments, the above method further comprises: Pre-process the complaint work order to obtain the pre-processed complaint work order; Determine the customer's emotional tendencies and keywords of the complaint incident based on the pre-processed complaint ticket; Based on the customer's emotional tendencies and keywords of the complaint incident, the urgency and importance of the complaint ticket are evaluated and the priority of the complaint ticket is determined.

[0049] Among them, emotional tendencies include anxiety, anger, sadness, and rage; keywords include entities, time, and other words.

[0050] For example, by analyzing the wording and tone in the pre-processed complaint ticket, we assessed that the customer's emotional tendency is relatively angry and anxious. Keywords include credit card and not received within a week, indicating that this complaint ticket has a certain degree of urgency and importance. It is determined to have a higher priority level so that it can be handled as soon as possible to avoid further deterioration of customer emotions.

[0051] Optionally, the order of each complaint work order is determined according to the category of the processing department of each complaint work order and the priority of each complaint work order, and each complaint work order is automatically assigned to the corresponding processing department in order.

[0052] Optionally, the status of complaint tickets is monitored and updated in real time to ensure that all complaint tickets are responded to within the specified time.

[0053] For example, when a complaint ticket flows from one processing link to the next, the status of the complaint ticket is updated in real time so that the processing progress of the complaint ticket can be viewed in real time, such as whether it has been assigned to a specific processing personnel, whether the processing personnel has started processing, and which stage the processing has reached, etc., so as to ensure that customer complaints are responded to within the specified time and avoid delays.

[0054] Step 130: Input the key features of the complaint ticket into the pre-trained second largest language model to obtain the complaint ticket processing strategy output by the second largest language model.

[0055] Among them, the second largest language model is trained based on the sample key features of the sample complaint tickets and the processing strategy labels of the sample complaint tickets.

[0056] Among them, the handling strategy for complaint tickets at least includes apology, compensation measures, prevention strategies and other handling decisions.

[0057] It should be noted that the second largest language model can generate personalized processing strategies based on the specific content and context of the complaint ticket.

[0058] It can be understood that by inputting the key features of the complaint ticket into the pre-trained second largest language model, and obtaining the complaint ticket processing strategy output by the second largest language model, not only can the efficiency of complaint handling be improved, but also it can ensure that all complaints are handled according to unified standards and procedures, which helps to maintain the consistency of service quality and reduce the instability of processing quality caused by individual differences.

[0059] In an embodiment of the present invention, by obtaining a complaint ticket to be processed, extracting the key features of the complaint ticket, inputting the key features of the complaint ticket into a pre-trained first language model, obtaining the processing department category of the complaint ticket output by the first language model, and inputting the key features of the complaint ticket into a pre-trained second language model, obtaining the processing strategy of the complaint ticket output by the second language model, not only can the processing departments of the complaint ticket be automatically classified, but also the processing strategy of the complaint ticket can be automatically generated, further improving the efficiency and quality of complaint handling.

[0060] Figure 2 A flow chart of inputting the key features of a complaint ticket into a pre-trained second language model is provided in an embodiment of the present invention. In some embodiments, the second language model includes an extraction layer, a matching layer, and a decision layer; Figure 2 As shown, in step 130, the key features of the complaint ticket are input into the pre-trained second language model to obtain the complaint ticket processing strategy output by the second language model, including: Step 131: Input the key features of the complaint ticket into the extraction layer to obtain the attribute features of each entity of the complaint ticket output by the extraction layer; Step 132: Input the attribute characteristics of each entity in the complaint ticket into the matching layer. The matching layer matches the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in the pre-built regulatory map. Based on the matching results, the guidance measures for the complaint ticket are determined. Step 133: Input the guidance measures of the complaint ticket to the decision-making layer, and obtain the handling strategy of the complaint ticket output by the decision-making layer. The handling strategy includes apology, compensation measures and prevention strategy.

[0061] Optionally, the extraction layer is used to extract the entities, attributes and attribute features of the complaint ticket; for example, the entity is a credit card, the attribute of the entity is the problem of the complaint, and the attribute feature is the specific feature of the problem.

[0062] Among them, the regulatory map is constructed based on multiple regulatory documents of the service industry. The regulatory map includes multiple entity nodes, corresponding attribute characteristics and guiding measures corresponding to each attribute.

[0063] Optionally, the matching layer matches each entity of the complaint ticket with each entity of the regulatory law map to obtain a matched target entity, and further matches the attribute characteristics of the target entity of the complaint ticket with the attribute characteristics of the target entity of the regulatory law map.

[0064] Among them, handling strategies include apologies, compensation measures (such as refunds, discounts, service upgrades, etc.) and prevention strategies (such as improving service processes, employee training, etc.).

[0065] Optionally, the decision-making layer generates a handling strategy for the complaint ticket by combining the guidance measures of the complaint ticket and historical case data.

[0066] In an embodiment of the present invention, through automated feature extraction, matching and decision-making processes, the need for manual intervention is greatly reduced, thereby speeding up complaint handling. By using a pre-built regulatory map for feature matching, it can not only ensure that each complaint is handled in accordance with unified standards, reducing human errors and biases, but also ensure that all handling measures comply with the requirements of relevant laws and regulations, reducing legal risks and potential fines caused by non-compliant operations. By generating preventive strategies, it can help enterprises take measures to prevent similar problems from happening again and continuously optimize business processes.

[0067] In some embodiments, the process of constructing a regulatory landscape includes: Build a regulatory knowledge base that stores multiple regulatory documents corresponding to complaint tickets; Extract the entity-relationship-entity in each regulatory document and construct an initial regulatory graph based on the entity-relationship-entity in each regulatory document; Clustering and screening the entities in the initial regulatory map to obtain an optimized regulatory map, and filling in the attributes of the entities in the optimized regulatory map to obtain a regulatory map; Determine the guiding measures corresponding to the attributes of each entity in the regulatory map based on multiple regulatory documents; Feature extraction is performed on the attributes of each entity in the regulatory law map to obtain the attribute features of each entity in the regulatory law map.

[0068] Alternatively, you can obtain relevant regulatory documents for the target industry by visiting the official websites of the State Administration for Market Regulation, local market supervision and administration bureaus, and industry associations, or by visiting the government's public information platforms and the information platforms of standard-setting organizations.

[0069] Optionally, each regulatory document is pre-processed by data cleaning, data conversion, word segmentation, etc. to obtain a pre-processed regulatory document, and feature extraction is performed on the pre-processed regulatory document to obtain multiple entity-relationship-entities, such as: credit card-business type-card application, credit card-business type-limit increase, etc.

[0070] Optionally, the entities in the initial regulatory law map are clustered, homogeneous entities are merged, and non-regulatory entities are eliminated.

[0071] Optionally, the attributes of each entity in the optimized regulatory map are filled in according to the contents of multiple regulatory documents. For example, card application complaints include question 1, question 2, and question 3, i.e., attribute 1, attribute 2, and attribute 3. Each attribute is filled in according to the content of each question.

[0072] Optionally, feature extraction is performed on the attributes of each entity in the regulatory law map so that attribute matching can be performed later based on the features matched on the complaint ticket to determine the attributes matched on the complaint ticket; for example, attribute 1 of the complaint of unissued card can be determined based on features A, B and C, and attribute 2 of the complaint of issued card can be determined based on features A, B and D.

[0073] Optionally, each entity node in the regulatory law map is tested to determine whether each entity node contains information on dimensions such as entities, relationships, attributes, and features, and nodes with missing content are marked to supplement or remove them.

[0074] It can be understood that by building a regulatory knowledge base, extracting the entity-relationship-entity in each regulatory document, constructing an initial regulatory map, optimizing and processing the initial regulatory map, and obtaining an information regulatory map containing dimensions such as entities, relationships, attributes, and characteristics, a comprehensive and rich data foundation is provided for subsequent decision-making processes, thereby improving the accuracy of decision-making.

[0075] In some embodiments, the attribute characteristics of each entity in the complaint ticket are matched with the attribute characteristics of each entity in the pre-built regulatory map, and the guiding measures for the complaint ticket are determined based on the matching results, including: The attribute characteristics of each entity in the complaint ticket are precisely matched with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, the guiding measures for the complaint ticket are determined based on the precise matching results. Otherwise, the attribute characteristics of each entity in the complaint ticket are fuzzy matched with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, the guiding measures for the complaint ticket are determined based on the fuzzy matching results. Otherwise, recommended guiding measures for the complaint ticket are determined.

[0076] Optionally, the attribute characteristics of each entity in the complaint ticket are matrixed to obtain a first matrix, the attribute characteristics of each entity in the regulatory regulations map are matrixed to obtain a second matrix, and the similarity between the first matrix and the second matrix is ​​calculated to obtain an exact matching result.

[0077] Optionally, fuzzy matching is performed between the attribute characteristics of each entity in the complaint ticket and the attribute characteristics of each entity in the regulatory regulations map to determine whether the match is successful, including: Calculate the similarity between the attribute characteristics of each entity in the complaint ticket and the attribute characteristics of each entity in the regulatory regulations map; Based on the similarity corresponding to the attribute features of each entity in the complaint ticket and the preset similarity threshold, it is determined whether there are matching attribute features. If so, the match is successful.

[0078] Optionally, in the event of a fuzzy match failure, the matching layer generates recommended guidance measures based on the key features of the complaint ticket and relevant historical cases, and identifies the source of the recommended guidance measures, such as "automatically generated by artificial intelligence AI."

[0079] It can be understood that by combining precise matching and fuzzy matching, the attribute characteristics of each entity in the complaint ticket are matched with the attribute characteristics of each entity in the pre-built regulatory map. Different guidance measures are output according to different matching results, which improves the accuracy of feature matching and ensures that corresponding guidance measures can be output in different matching situations.

[0080] In some embodiments, the above method further comprises: Obtain feedback data on the handling strategy of complaint tickets; Feedback data on the complaint ticket handling strategy is input into the evaluation layer of the pre-trained third language model, and the evaluation layer outputs a quality assessment result of the complaint ticket handling strategy. Based on the quality assessment results of the complaint ticket handling strategy, the parameters of the second largest language model are optimized; Among them, the evaluation layer of the third language model is trained based on historical feedback data of historical processing strategies of historical complaint tickets.

[0081] Optionally, after the target processing department processes the complaint ticket according to the complaint ticket processing strategy, feedback data of the complaint ticket processing strategy is obtained, such as customer satisfaction, service quality evaluation, processing speed and other data.

[0082] Optionally, the evaluation criteria of the assessment layer include whether the processing result is satisfactory, whether the processing speed is fast, whether repeated processing occurs, etc.

[0083] For example, after implementing the handling strategy for customer Mr. Sun's complaint ticket, we collect Mr. Sun's feedback and evaluate whether he is satisfied with the handling results and whether the handling speed is fast, and obtain a quality score for the handling strategy. If Mr. Sun's feedback is satisfactory and the handling speed is fast, the quality score will be higher.

[0084] Optionally, when the quality score of the complaint ticket processing strategy is lower than a preset threshold, improvement suggestions are obtained based on the feedback data, such as further improving the processing speed or adding compensatory measures, and the parameters of the second largest language model are optimized according to the improvement suggestions.

[0085] Optionally, an online learning algorithm is applied to enable the second largest language model to self-optimize according to the post-evaluation results, thereby improving the accuracy and compliance of the processing strategy.

[0086] For example, if it is found that the customer is relatively satisfied with a certain compensation method, then this compensation method can be given priority in the subsequent processing of similar complaint tickets.

[0087] In some embodiments, the training process of the evaluation layer of the third language model includes: Initialize the evaluation layer of the third language model to obtain the initial evaluation layer; Obtain historical feedback data on historical handling strategies for historical complaint tickets; Determine the quality assessment result label of the historical handling strategy of historical complaint tickets; Input the historical feedback data of the historical handling strategies of historical complaint work orders into the initial evaluation layer, and obtain the quality evaluation results of the historical handling strategies of historical complaint work orders output by the initial evaluation layer; Based on the quality evaluation results of historical processing strategies of historical complaint work orders and the quality evaluation result labels of historical processing strategies of historical complaint work orders, the parameters of the initial evaluation layer are iteratively optimized to obtain the evaluation layer.

[0088] It can be understood that by obtaining feedback data on the handling strategy of complaint tickets, the feedback data is evaluated using the evaluation layer of the pre-trained third language model to obtain the quality evaluation results of the handling strategy, and then the parameters of the second language model can be optimized based on the quality evaluation results, which can ensure that the handling strategy is more in line with user expectations and regulatory requirements, thereby improving the overall service quality and enhancing the user experience.

[0089] In some embodiments, the above method further comprises: The complaint ticket handling strategy is input into the decision layer of the pre-trained third language model. The decision layer determines whether the handling strategy meets the regulatory exclusion scenario and obtains the decision result of the complaint ticket handling strategy. If it is determined that the handling strategy of the complaint ticket meets the regulatory exclusion scenario, the handling strategy of the complaint ticket will be eliminated; Optimize the parameters of the second language model based on the results of the complaint ticket processing strategy; Among them, the judgment layer of the third language model is trained based on the sample processing strategies of sample complaint tickets and the judgment result labels of the sample processing strategies.

[0090] Among them, regulatory exclusion scenarios are determined based on multiple regulatory documents. For example, for claim delays caused by "force majeure", the corresponding complaint tickets and their handling strategies can be excluded.

[0091] For example, a customer, Ms. Zhou, complained that the insurance claim was processed too slowly. Through the classification algorithm, analysis found that Ms. Zhou's complaint was due to delays in claims due to force majeure factors such as natural disasters. The complaint was then eliminated and no further processing was performed.

[0092] In some embodiments, the training process of the decision layer of the third language model includes: Initialize the decision layer of the third language model to obtain the initial decision layer; Obtain sample handling strategies for sample complaint tickets; Determine the judgment result label of the sample handling strategy of the sample complaint work order; The sample handling strategy of the sample complaint work order is input into the initial judgment layer, and the initial judgment layer determines whether the sample handling strategy of the sample complaint work order meets the regulatory elimination scenario, and obtains the judgment result of the sample handling strategy of the sample complaint work order; Based on the determination results of the sample processing strategies of the sample complaint work orders and the determination result labels of the sample processing strategies of the sample complaint work orders, the parameters of the initial determination layer are iteratively optimized to obtain the determination layer.

[0093] It can be understood that by utilizing the judgment layer of the third language model, it is possible to determine whether the handling strategy of the complaint ticket complies with the regulatory exclusion scenario, and then the handling strategy that complies with the regulatory exclusion scenario can be eliminated, and the parameters of the second language model can be optimized, thereby improving the performance of the second language model and enabling it to output a handling strategy with greater rationality and compliance.

[0094] Complaint Case: Customer Ms. Zhang complained that the bank’s credit card overdue fees were too high, believing that the bank’s failure to remind her in a timely manner resulted in high overdue fees on her credit card.

[0095] The complaint handling process includes the following steps: Step 1: Organize the complaint incidents; Specifically, through voice recognition, Ms. Zhang's complaint call was converted into a complaint work order, and key information was extracted, such as the customer's identity was Ms. Zhang, the reason for the complaint was the high overdue fees for credit cards, and the request to reduce the overdue fees or waive some fees; at the same time, it was assessed that Ms. Zhang was emotionally agitated and dissatisfied; further complaint features were extracted to determine that Ms. Zhang had a good credit record in the past, and this overdue payment may have been due to forgetting to repay under special circumstances.

[0096] Step 2: Work order transfer; Specifically, based on the characteristics of the complaint, the department that handles the complaint ticket is determined to be the credit card business processing department, and the complaint ticket is assigned to this department; the emotional tendencies and keywords in the complaint ticket are analyzed to determine that the complaint ticket has a certain degree of urgency and importance, and the processing priority is increased; the status of the complaint ticket is monitored in real time through an event-driven architecture to ensure that the processing progress can be tracked.

[0097] Step 3: Generate processing strategy; Specifically, the system searches the knowledge base for relevant regulatory requirements to ensure compliance with the handling strategy. For example, it checks whether the bank's overdue fees comply with regulatory requirements. By referencing historical similar cases and examining other cases where fees were waived due to special circumstances, it provides customer service staff with handling suggestions and automatically generates a handling strategy. For example, "Dear Ms. Zhang, we understand that your credit card overdue payment may be due to special circumstances. After verification, we have decided to waive some of your overdue fees. At the same time, we will strengthen repayment reminders to prevent similar situations from happening again."

[0098] Step 4: Post-evaluation of treatment strategies; Specifically, the complaint case was evaluated to see whether it met the elimination scenario, which was obviously not the case. After the processing strategy was implemented, Ms. Zhang's feedback was collected. If Ms. Zhang was satisfied, a higher score would be given. If she was not satisfied, improvement suggestions would be given based on the feedback.

[0099] Figure 3A flow chart of the training process of the second language model provided in an embodiment of the present invention is shown as follows: Figure 3 As shown, in some embodiments, the training process of the second largest language model includes: Step 310: Obtain a sample complaint work order and extract key features of the sample complaint work order; Step 320: Determine the processing strategy label of the sample complaint work order; Step 330: Input the sample key features of the sample complaint worksheet into the initial second largest language model to obtain the processing strategy of the sample complaint worksheet output by the initial second largest language model; Step 340: Based on the processing strategy of the sample complaint ticket and the processing strategy label of the sample complaint ticket, calculate the loss function value, iteratively optimize the parameters of the initial second largest language model, and obtain the second largest language model.

[0100] Optionally, historical complaint case data is obtained, pre-processed to obtain sample complaint work orders, and processing strategy labels for the sample complaint work orders are determined.

[0101] Optionally, extract sample key features of the sample complaint ticket, including: Perform preprocessing on the sample complaint ticket, including word segmentation, part-of-speech tagging, and entity recognition, to obtain the preprocessed sample complaint ticket; The preprocessed sample complaint work order is input into the pre-trained fourth language model to obtain the sample key features of the sample complaint work order output by the fourth language model.

[0102] Optionally, the initial second largest language model includes an initial extraction layer, an initial matching layer, and an initial decision layer; Correspondingly, the sample key features of the sample complaint ticket are input into the initial second-largest language model to obtain the processing strategy of the sample complaint ticket output by the initial second-largest language model, including: Input the sample key features of the sample complaint work order into the initial extraction layer to obtain the sample attribute features of each sample entity of the sample complaint work order output by the initial extraction layer; Input the sample attribute characteristics of each sample entity of the sample complaint work order into the initial matching layer, which matches the sample attribute characteristics of each sample entity of the sample complaint work order with the attribute characteristics of each entity in the pre-constructed regulatory law map. Based on the sample matching results, the sample guidance measures for the sample complaint work order are determined; The sample guidance measures of the sample complaint work order are input into the initial decision-making layer, and the processing strategy of the sample complaint work order output by the initial decision-making layer is obtained. The processing strategy includes apology, compensation measures and prevention strategy.

[0103] Optionally, the sample attribute characteristics of each sample entity of the sample complaint work order are matched with the attribute characteristics of each entity of the pre-built regulatory law map. Based on the sample matching results, sample guidance measures for the sample complaint work order are determined, including: The sample attribute characteristics of each sample entity of the sample complaint work order are precisely matched with the attribute characteristics of each entity of the regulatory law map to determine whether the match is successful. If so, the sample guidance measures of the sample complaint work order are determined based on the sample precise matching results. Otherwise, the sample attribute characteristics of each sample entity of the sample complaint work order are fuzzy matched with the attribute characteristics of each entity of the regulatory law map to determine whether the match is successful. If so, the sample guidance measures of the sample complaint work order are determined based on the sample fuzzy matching results. Otherwise, the recommended sample guidance measures for the sample complaint work order are determined.

[0104] It can be understood that by using the sample key features of the sample complaint tickets as training samples, the processing strategy labels of the sample complaint tickets as sample labels, training the initial second-largest language model, and iteratively optimizing the parameters of the second-largest language model, the performance of the second-largest language model can be improved, enabling it to more accurately predict the processing strategies of complaint tickets and provide suggestions that are more in line with actual conditions and regulatory regulations.

[0105] In some embodiments, after obtaining the second largest language model, the method further includes: Obtain feedback data on the handling strategies of sample complaint tickets; Feedback data on the sample complaint ticket handling strategy is input into the evaluation layer of the pre-trained third language model, and the evaluation layer outputs a quality evaluation result of the sample complaint ticket handling strategy. The sample complaint ticket's handling strategy is input into the decision layer of the pre-trained third language model. The decision layer determines whether the sample complaint ticket's handling strategy meets the regulatory exclusion scenario, and obtains the decision result of the sample complaint ticket's handling strategy. Based on the quality assessment results of the sample complaint ticket processing strategy and the determination results of the sample complaint ticket processing strategy, the parameters of the second largest language model are optimized.

[0106] Optionally, based on the quality scores of the processing strategies given by the third language model and a preset score threshold, high-quality processing strategies (such as processing strategies with scores above 80) are screened out and included in the case library as positive samples for learning by the second language model; low-quality processing strategies (such as processing strategies with scores below 60) are screened out and included in the case library as negative samples for learning by the second language model.

[0107] Optionally, the second language model and the third language model are trained against each other, and through iterative feedback, cross-validation, and optimization, the generalization ability and accuracy of the second language model and the third language model can be improved.

[0108] Optionally, a series of financial regulatory exclusion provisions are collated, and relevant cases are input into the third language model for targeted training to optimize the third language model's ability to judge complaint regulatory exclusion provisions.

[0109] Optionally, the third language model makes an intelligent judgment on the processing strategy generated by the second language model to determine whether it complies with the regulatory exclusion scenario. If it does, the corresponding exclusion regulations are provided; if it does not, it is forwarded.

[0110] Optionally, the second largest language model optimizes the processing strategy based on the elimination criteria fed back by the third largest language model, and the third largest language model again judges the optimized processing strategy to determine whether it meets the regulatory elimination scenario, and the cycle is repeated. If it fails in all three rounds (i.e., it meets the elimination scenario), the corresponding elimination criteria will be included in the corpus of the second largest language model for optimization; if it passes within three rounds (i.e., it does not meet the elimination scenario), the corresponding elimination criteria will be included in the corpus of the third largest language model for optimization.

[0111] It can be understood that by inputting the processing strategy output by the second largest language model into the third largest language model for post-evaluation, the quality assessment results and judgment results of the processing strategy are obtained, and then the second largest language model can be optimized based on the quality assessment results and judgment results, thereby improving the second largest language model's prediction accuracy for the processing strategy.

[0112] The following describes a complaint ticket processing device provided in an embodiment of the present invention. The device described below and the complaint ticket processing method described above can be referenced to each other.

[0113] Figure 4 A schematic diagram of the structure of the complaint work order processing device provided by an embodiment of the present invention is shown as follows: Figure 4 As shown, the complaint work order processing device 400 includes: A feature extraction unit 410 is used to obtain a complaint ticket to be processed and extract key features of the complaint ticket; A classification unit 420 is configured to input the key features of the complaint ticket into a pre-trained first language model to obtain a processing department category of the complaint ticket output by the first language model; A decision unit 430 is configured to input the key features of the complaint ticket into a pre-trained second language model to obtain a handling strategy for the complaint ticket output by the second language model; Among them, the first language model is trained based on the sample key features of the sample complaint work orders and the processing department category labels of the sample complaint work orders. The second language model is trained based on the sample key features of the sample complaint work orders and the processing strategy labels of the sample complaint work orders.

[0114] Optionally, the second language model includes an extraction layer, a matching layer, and a decision layer; Correspondingly, the key features of the complaint ticket are input into the pre-trained second largest language model to obtain the complaint ticket processing strategy output by the second largest language model, including: Input the key features of the complaint ticket into the extraction layer, and obtain the attribute features of each entity of the complaint ticket output by the extraction layer; The attribute characteristics of each entity in the complaint ticket are input into the matching layer, which matches the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in the pre-built regulatory map. Based on the matching results, the guiding measures for the complaint ticket are determined; The guiding measures of the complaint work order are input into the decision-making layer, and the decision-making layer outputs the handling strategy of the complaint work order, which includes apology, compensation measures and prevention strategy.

[0115] Optionally, the process of building a regulatory map includes: Build a regulatory knowledge base that stores multiple regulatory documents corresponding to complaint tickets; Extract the entity-relationship-entity in each regulatory document and construct an initial regulatory graph based on the entity-relationship-entity in each regulatory document; Clustering and screening the entities in the initial regulatory map to obtain an optimized regulatory map, and filling in the attributes of the entities in the optimized regulatory map to obtain a regulatory map; Determine the guiding measures corresponding to the attributes of each entity in the regulatory map based on multiple regulatory documents; Feature extraction is performed on the attributes of each entity in the regulatory law map to obtain the attribute features of each entity in the regulatory law map.

[0116] Optionally, the attribute characteristics of each entity in the complaint ticket are matched with the attribute characteristics of each entity in the pre-built regulatory map. Based on the matching results, the guiding measures for the complaint ticket are determined, including: The attribute characteristics of each entity in the complaint ticket are precisely matched with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, the guiding measures for the complaint ticket are determined based on the precise matching results. Otherwise, the attribute characteristics of each entity in the complaint ticket are fuzzy matched with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, the guiding measures for the complaint ticket are determined based on the fuzzy matching results. Otherwise, recommended guiding measures for the complaint ticket are determined.

[0117] Optionally, the complaint ticket processing device 400 further includes: An acquisition unit, used to obtain feedback data on the handling strategy of the complaint ticket; An evaluation unit, configured to input feedback data on the complaint ticket handling strategy into the evaluation layer of the pre-trained third language model, and obtain a quality evaluation result of the complaint ticket handling strategy output by the evaluation layer; The first optimization unit is used to optimize the parameters of the second language model based on the quality evaluation results of the complaint ticket processing strategy; Among them, the evaluation layer of the third language model is trained based on historical feedback data of historical processing strategies of historical complaint tickets.

[0118] In some embodiments, the complaint ticket processing apparatus 400 further includes: A determination unit, which is used to input the complaint ticket handling strategy into the determination layer of the pre-trained third language model. The determination layer determines whether the handling strategy meets the regulatory exclusion scenario and obtains the determination result of the complaint ticket handling strategy; The rejection unit is used to reject the handling strategy of the complaint ticket if it is determined that the handling strategy of the complaint ticket meets the regulatory rejection scenario; A second optimization unit is used to optimize the parameters of the second language model according to the determination result of the handling strategy of the complaint ticket; Among them, the judgment layer of the third language model is trained based on the sample processing strategies of sample complaint tickets and the judgment result labels of the sample processing strategies.

[0119] Optionally, the training process of the second language model includes: Obtain sample complaint tickets and extract key features of the sample complaint tickets; Determine the handling strategy label for the sample complaint ticket; Input the sample key features of the sample complaint work order into the initial second largest language model to obtain the processing strategy of the sample complaint work order output by the initial second largest language model; Based on the processing strategy of the sample complaint work order and the processing strategy label of the sample complaint work order, the loss function value is calculated, and the parameters of the initial second-largest language model are iteratively optimized to obtain the second-largest language model.

[0120] Optionally, the training process of the second language model further includes: Obtain feedback data on the handling strategies of sample complaint tickets; Feedback data on the sample complaint ticket handling strategy is input into the evaluation layer of the pre-trained third language model, and the evaluation layer outputs a quality evaluation result of the sample complaint ticket handling strategy. The sample complaint ticket's handling strategy is input into the decision layer of the pre-trained third language model. The decision layer determines whether the sample complaint ticket's handling strategy meets the regulatory exclusion scenario, and obtains the decision result of the sample complaint ticket's handling strategy. Based on the quality assessment results of the sample complaint ticket processing strategy and the determination results of the sample complaint ticket processing strategy, the parameters of the second largest language model are optimized.

[0121] It should be noted here that the complaint ticket processing device provided in the embodiment of the present invention can implement all the method steps implemented in the above-mentioned complaint ticket processing method embodiment, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as the method embodiment will not be described in detail here.

[0122] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute a method for processing a complaint ticket, which includes: obtaining a complaint ticket to be processed and extracting key features of the complaint ticket; inputting the key features of the complaint ticket into a pre-trained first language model to obtain a processing department category of the complaint ticket output by the first language model; inputting the key features of the complaint ticket into a pre-trained second language model to obtain a processing strategy for the complaint ticket output by the second language model; wherein the first language model is trained based on sample key features of sample complaint tickets and the processing department category labels of the sample complaint tickets; and the second language model is trained based on sample key features of sample complaint tickets and the processing strategy labels of the sample complaint tickets.

[0123] Furthermore, the logic instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0124] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the complaint ticket processing method provided by the above-mentioned methods, the method comprising: obtaining a complaint ticket to be processed and extracting key features of the complaint ticket; inputting the key features of the complaint ticket into a pre-trained first language model to obtain the processing department category of the complaint ticket output by the first language model; inputting the key features of the complaint ticket into a pre-trained second language model to obtain the processing strategy of the complaint ticket output by the second language model; wherein the first language model is trained based on the sample key features of the sample complaint ticket and the processing department category label of the sample complaint ticket; the second language model is trained based on the sample key features of the sample complaint ticket and the processing strategy label of the sample complaint ticket.

[0125] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0126] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for processing a complaint work order, characterized in that: include: Obtain a complaint ticket to be processed, and extract key features of the complaint ticket; Inputting the key features of the complaint worksheet into a pre-trained first language model to obtain the processing department category of the complaint worksheet output by the first language model; Inputting key features of the complaint ticket into a pre-trained second largest language model to obtain a processing strategy for the complaint ticket output by the second largest language model; Among them, the first largest language model is trained based on the sample key features of the sample complaint worksheets and the processing department category labels of the sample complaint worksheets; the second largest language model is trained based on the sample key features of the sample complaint worksheets and the processing strategy labels of the sample complaint worksheets.

2. The method for handling complaint tickets according to claim 1, characterized in that: The second largest language model includes an extraction layer, a matching layer, and a decision layer; Correspondingly, the key features of the complaint ticket are input into the pre-trained second largest language model to obtain the processing strategy of the complaint ticket output by the second largest language model, including: Inputting the key features of the complaint worksheet into the extraction layer, and obtaining the attribute features of each entity of the complaint worksheet output by the extraction layer; Input the attribute characteristics of each entity of the complaint worksheet into the matching layer, and the matching layer matches the attribute characteristics of each entity of the complaint worksheet with the attribute characteristics of each entity of the pre-constructed regulatory law map, and determines the guiding measures of the complaint worksheet according to the matching results; The guiding measures of the complaint worksheet are input into the decision-making layer, and the processing strategy of the complaint worksheet output by the decision-making layer is obtained, and the processing strategy includes apology, compensation measures and prevention strategy.

3. The method for handling complaint work orders according to claim 2, characterized in that: The construction process of the regulatory law map includes: Building a regulatory knowledge base, wherein the regulatory knowledge base stores multiple regulatory documents corresponding to the complaint work order; Extracting the entity-relationship-entity in each regulatory document, and constructing an initial regulatory map based on the entity-relationship-entity in each regulatory document; Clustering and screening each entity in the initial regulatory law map to obtain an optimized regulatory law map, and filling in the attributes of each entity in the optimized regulatory law map to obtain the regulatory law map; Determining, based on the multiple regulatory documents, guiding measures corresponding to the attributes of each entity in the regulatory map; Feature extraction is performed on the attributes of each entity in the regulatory law map to obtain attribute features of each entity in the regulatory law map.

4. The method for handling complaint work orders according to claim 2, characterized in that: The method of matching the attribute characteristics of each entity of the complaint worksheet with the attribute characteristics of each entity of the pre-constructed regulatory law map, and determining the guiding measures for the complaint worksheet according to the matching results, includes: Precisely match the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, determine the guidance measures for the complaint ticket based on the precise matching result. Otherwise, fuzzy match the attribute characteristics of each entity in the complaint ticket with the attribute characteristics of each entity in the regulatory law map to determine whether the match is successful. If so, determine the guidance measures for the complaint ticket based on the fuzzy matching result. Otherwise, determine the recommended guidance measures for the complaint ticket.

5. The method for handling complaint tickets according to claim 1, characterized in that: The method further comprises: Obtaining feedback data on the handling strategy of the complaint ticket; Inputting feedback data of the handling strategy of the complaint ticket into the evaluation layer of the pre-trained third language model to obtain a quality evaluation result of the handling strategy of the complaint ticket output by the evaluation layer; Optimizing the parameters of the second largest language model according to the quality evaluation result of the handling strategy of the complaint ticket; Among them, the evaluation layer of the third language model is trained based on historical feedback data of historical processing strategies of historical complaint work orders.

6. The method for handling complaint work orders according to claim 1, characterized in that: The method further comprises: Input the processing strategy of the complaint ticket into the decision layer of the pre-trained third language model, and the decision layer determines whether the processing strategy meets the regulatory rejection scenario, thereby obtaining a decision result of the processing strategy of the complaint ticket; If it is determined that the handling strategy of the complaint ticket meets the regulatory elimination scenario, eliminate the handling strategy of the complaint ticket; Optimizing the parameters of the second largest language model according to the determination result of the processing strategy of the complaint ticket; Among them, the decision layer of the third language model is trained based on the sample processing strategy of the sample complaint work order and the decision result label of the sample processing strategy.

7. The method for handling complaint work orders according to claim 1, characterized in that: The training process of the second largest language model includes: Obtain a sample complaint work order, and extract sample key features of the sample complaint work order; Determine a handling strategy label for the sample complaint ticket; Inputting the sample key features of the sample complaint worksheet into the initial second largest language model to obtain a processing strategy for the sample complaint worksheet output by the initial second largest language model; Based on the processing strategy of the sample complaint work order and the processing strategy label of the sample complaint work order, the loss function value is calculated, and the parameters of the initial second largest language model are iteratively optimized to obtain the second largest language model.

8. The method for handling complaint work orders according to claim 7, characterized in that: After obtaining the second largest language model, the method further includes: Obtaining feedback data on the processing strategy of the sample complaint work order; Inputting feedback data of the processing strategy of the sample complaint worksheet into the evaluation layer of the pre-trained third language model to obtain a quality evaluation result of the processing strategy of the sample complaint worksheet output by the evaluation layer; Inputting the processing strategy of the sample complaint work order into the decision layer of the pre-trained third language model, the decision layer determines whether the processing strategy of the sample complaint work order meets the regulatory exclusion scenario, and obtains the decision result of the processing strategy of the sample complaint work order; The parameters of the second largest language model are optimized according to the quality evaluation result of the processing strategy of the sample complaint worksheet and the determination result of the processing strategy of the sample complaint worksheet.

9. A complaint work order processing device, characterized in that: include: A feature extraction unit, used to obtain a complaint ticket to be processed and extract key features of the complaint ticket; A classification unit, used for inputting key features of the complaint ticket into a pre-trained first language model to obtain a processing department category of the complaint ticket output by the first language model; A decision-making unit, used for inputting key features of the complaint ticket into a pre-trained second largest language model to obtain a processing strategy for the complaint ticket output by the second largest language model; Among them, the first largest language model is trained based on the sample key features of the sample complaint worksheets and the processing department category labels of the sample complaint worksheets, and the second largest language model is trained based on the sample key features of the sample complaint worksheets and the processing strategy labels of the sample complaint worksheets.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the complaint ticket processing method as described in any one of claims 1 to 8 is implemented.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the complaint ticket processing method as described in any one of claims 1 to 8.

Images