Multi-intent scene customer service script processing method and device, equipment and medium

Abstract

The application relates to the technical field of intelligent customer service and can be applied to the fields of finance and medicine, and discloses a multi-intent scene customer service dialogue processing method, device, equipment and medium. After obtaining user input data, an intent analysis result is obtained through multi-dimensional intent analysis. According to the intent analysis result, a candidate dialogue set is recalled in the question and answer knowledge base. When the intent analysis result and the candidate dialogue set meet preset interception conditions, the pushing process of the standard reply dialogue is intercepted, and rewritten prompt information is obtained. Then, a large language model is used to perform dynamic rewriting based on the rewritten prompt information to obtain a rewritten reply dialogue, which is pushed to a user terminal. The method can reconstruct a dialogue output link through a dialogue fusion interception mechanism, and can compress multiple highly coincident static dialogues into a single refined reply through semantic deduplication and information dimension reduction of the large language model, thereby reducing the transmission rate of redundant information and improving the reply redundancy problem in the multi-intent scene.

Description

Technical Field

[0001] This application relates to the field of intelligent customer service technology, and in particular to a method, device, equipment and medium for processing customer service scripts in multi-intent scenarios. Background Technology

[0002] Intelligent customer service is a system that uses artificial intelligence technology to simulate the thinking and behavior of customer service personnel, providing automated and intelligent services to customers. Intelligent customer service can integrate technologies such as Natural Language Processing (NLP), machine learning, knowledge graphs, and speech recognition. By understanding the questions entered by customers and automatically matching answers from a knowledge base, it can complete business transactions, provide business consultations, and even perceive emotions.

[0003] In business areas such as insurance claims, financial services, and medical services, intelligent customer service requires the collection of large amounts of customer information and documents. It employs a dialogue interaction paradigm based on intent recognition and static mapping of standard responses. A natural language understanding module parses the user's voice or text input, extracting key entities and business intents. Then, a dialogue management module searches a pre-configured static knowledge base or response template library based on the extracted intent, matching and recalling the corresponding standard response, which is then output to the user.

[0004] Because user input sentences may be highly condensed and contain multiple directions, the dialogue architecture of intelligent customer service will trigger multiple independent skill branches when processing such multi-intent input sentences. This causes the intelligent customer service system to recall a static script for each intent and push these semantically highly overlapping redundant responses to the user simultaneously or continuously, resulting in a "mechanical screen-swiping" problem in multi-intent concurrent scenarios, which reduces the user's interactive experience. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method, apparatus, device, and medium for processing customer service scripts in multi-intent scenarios, in order to solve the problem of redundant responses in intelligent customer service in multi-intent scenarios.

[0006] According to a first aspect of this application, a method for processing customer service scripts in multi-intent scenarios is provided, the method comprising: Acquire user input data, wherein the user input data is based on data input from the user terminal; Perform multidimensional intent parsing on the user input data to obtain intent parsing results that include at least one input intent; Based on the intent parsing result, a candidate set of dialogues is recalled from the question-and-answer knowledge base, the candidate set of dialogues including at least one standard response dialogue that matches the input intent; When the intent parsing result and the candidate dialogue set meet the preset interception conditions, a rewrite prompt message is obtained; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt message includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script; The rewritten response message is pushed to the user's terminal.

[0007] According to a second aspect of this application, a multi-intent scenario customer service script processing device is provided, the device comprising: The data input module is used to acquire user input data, which is data input based on the user terminal. A multidimensional intent parsing module is used to perform multidimensional intent parsing on the user input data to obtain an intent parsing result that includes at least one input intent; The script retrieval module is used to retrieve a set of candidate scripts from the question-and-answer knowledge base based on the intent parsing result. The set of candidate scripts includes at least one standard response script that matches the input intent. An interception decision module is used to obtain rewrite prompt information when the intent parsing result and the candidate dialogue set meet preset interception conditions; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt information includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; The perceptual rewriting module is used to perform dynamic rewriting based on the rewriting prompt information using a large language model to obtain a rewritten response script; The script push module is used to push the rewritten reply script to the user terminal.

[0008] According to a third aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described multi-intent scenario customer service script processing method.

[0009] According to a fourth aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described multi-intent scenario customer service script processing method.

[0010] The aforementioned multi-intent scenario customer service script processing method involves performing multi-dimensional intent parsing on user input data after acquisition to obtain intent parsing results including at least one input intent. Based on the intent parsing results, a candidate script set including at least one standard response script is retrieved from the question-and-answer knowledge base. When the intent parsing results and the candidate script set meet preset interception conditions, the push process of the standard response script is intercepted, and rewriting prompt information is obtained. Then, a large language model is used to perform dynamic rewriting based on the rewriting prompt information to obtain a rewritten response script, which is then pushed to the user terminal. This method can reconstruct the dialogue output link through a script fusion interception mechanism and compress multiple highly overlapping static scripts into a single concise response through semantic deduplication and information dimensionality reduction using a large language model, reducing the transmission rate of redundant information and improving the response redundancy problem in multi-intent scenarios.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the application environment provided for the embodiments of this application; Figure 2 This is a flowchart illustrating a multi-intent scenario customer service script processing method provided in an embodiment of this application. Figure 3 This is a flowchart of the multi-intent speech dynamic fusion processing provided in the embodiments of this application; Figure 4 This application provides an architecture diagram of an intelligent dialogue system. Figure 5 This application provides a sequence diagram of multi-intent scenario interaction. Figure 6 This is a flowchart of semantic deduplication and information dimensionality reduction processing provided in the embodiments of this application; Figure 7 A schematic diagram of the guardrail mechanism and illusion control strategy provided in the embodiments of this application; Figure 8This is a schematic diagram of the structure of the multi-intent scenario customer service script processing device provided in the embodiments of this application. Detailed Implementation

[0014] 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, 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.

[0015] In this embodiment, the multi-intent scenario customer service script processing method, apparatus, device, and medium can be applied to intelligent customer service platforms in fields such as finance, healthcare, and insurance. Intelligent customer service platforms can replace customer service personnel in providing customer service. Specifically, intelligent customer service is a system that uses artificial intelligence technology to simulate the thinking and behavior of customer service personnel, providing automated and intelligent services to customers.

[0016] An intelligent customer service platform, as a type of intelligent question answering system (IQAS), is an artificial intelligence application that can understand questions posed by users in natural language and automatically retrieve, extract, and generate accurate answers from knowledge bases or data sources. For example, an intelligent customer service platform can integrate technologies such as Natural Language Processing (NLP), machine learning, knowledge graphs, and speech recognition. By understanding the questions entered by customers and automatically matching answers from the knowledge base, it can complete business processing, provide business consultations, and even handle emotion perception.

[0017] like Figure 1 As shown, an intelligent customer service platform can include a client and a server. The client refers to the terminal through which users directly interact, responsible for data collection, display, and initial verification, and for sending user requests to the server. The client can also include applications, web pages, and operating systems deployed on the terminal device.

[0018] Depending on the business scenario corresponding to the application environment, clients can be of different types. For example, in a financial insurance claims scenario, clients can be divided into user clients and operation and maintenance management clients based on the user's target audience. User clients can be accessed on the user's terminal device through applications, mini-programs, web pages, etc. Operation and maintenance management clients can be accessed on the terminal devices used by operation and maintenance personnel through a backend management system (web) or a monitoring dashboard for operation and maintenance interactions.

[0019] When the application's business scenario is a healthcare service scenario, the client can be categorized based on the user: patient-side, doctor-side, operations and maintenance management-side, and integration-side. The patient-side allows user interaction via applications, mini-programs, or web pages on the terminal devices used by patients or related users (such as family members). The doctor-side allows user interaction via desktop applications, mobile workstation apps, or dedicated tablet interfaces on the terminal devices used by doctors or other healthcare-related users. The operations and maintenance management-side allows interaction via a backend management system (web) or monitoring dashboard on the terminal devices used by operations and maintenance personnel. The integration-side generally refers to terminal devices such as Hospital Information System (HIS) interfaces, medical insurance settlement terminals, and wearable device gateways, used for interactions related to performing integration functions.

[0020] Servers can be deployed in the cloud or data centers, responsible for core business processing, persistent data storage, security management, and integration with external systems. In some embodiments, the server may include one or more combinations of servers such as web servers, application servers, database servers, and storage servers.

[0021] The client and server can establish a communication connection via a network, thereby building an intelligent customer service platform. For example, when the intelligent customer service platform is applied to business scenarios such as finance and insurance, it can perform intelligent question-and-answer interaction based on the user's natural language text in the business scenario. By combining the user's question information with relevant knowledge retrieval based on business data generated in the business scenario, it can generate and provide feedback on financial and insurance business-related answers to meet business needs such as business consultation, business processing, and business recommendations.

[0022] In some embodiments, in business scenarios such as insurance claims, financial services, and medical services, where intelligent customer service requires the collection of large amounts of customer information and documents, a dialogue interaction paradigm based on intent recognition and static mapping of standard dialogue scripts is adopted. A natural language understanding module parses the user's voice or text input to extract key entities and business intents from the input content. Then, a dialogue management module searches a pre-configured static knowledge base or dialogue script template library based on the extracted intent, matches and recalls the corresponding standard response script, and outputs the recalled response script to the user.

[0023] Because user input sentences may be highly condensed and contain multiple directions, the dialogue architecture of intelligent customer service will trigger multiple independent skill branches when processing such multi-intent input sentences. This causes the intelligent customer service system to recall a static script for each intent and push these semantically highly overlapping redundant responses to the user simultaneously or continuously, resulting in a "mechanical screen-swiping" problem in multi-intent concurrent scenarios, which reduces the user's interactive experience.

[0024] To address the issue of redundant responses in multi-intent scenarios for intelligent customer service, this application provides a method for processing customer service scripts in multi-intent scenarios, applicable to the financial and medical fields. This method can reconstruct the dialogue output link through a script fusion interception mechanism, and compress multiple highly overlapping static scripts into a single concise response through semantic deduplication and information dimensionality reduction using a large language model. This reduces the transmission rate of redundant information and improves the response redundancy problem in multi-intent scenarios.

[0025] The method can be applied to intelligent customer service platforms or electronic devices that establish a communication connection with the intelligent customer service platform and have data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control machines. For ease of description, the intelligent customer service platform is used as the execution subject of the method in this embodiment. It should be understood that the method can also be applied to other types of execution subjects, which are not illustrated in this embodiment. Figure 2 As shown, the method includes: S101. Obtain user input data.

[0026] In the process of providing intelligent customer service, the intelligent customer service platform can first obtain user input data. The user data refers to one or more data sets input by the user through the user terminal. For example, the user input data includes the original input text based on the user terminal input, or text data obtained by converting other types of user input data through data format conversion.

[0027] The original input text can be text data in the form of natural language text. For example, when a user interacts through an intelligent customer service platform, they can directly input text through their user terminal. In this case, the original input text can be obtained by receiving the text data uploaded by the user terminal.

[0028] When obtaining text data through data format conversion, built-in data format conversion algorithms such as speech-to-text and image-to-text can be used to convert other types of data into text data. For example, users can input audio data via voice input. For audio data input via voice, the voice-to-text tool built into the user terminal or intelligent customer service platform can be used to convert the audio data into text data, obtaining the original input text.

[0029] In some embodiments, in addition to the original input text, user input data may also include other data related to the customer service dialogue process, such as user identification information, user-uploaded data, historical dialogue information, and historical context information of the current dialogue.

[0030] S102. Perform multidimensional intent parsing on the user input data to obtain intent parsing results that include at least one input intent.

[0031] After obtaining user input data, text data can be extracted from the user input data, and multidimensional intent parsing can be performed on the text data to obtain intent parsing results. The intent parsing results include at least one input intent.

[0032] Input intent refers to an atomic business goal or service request achieved through a single interaction within the current business scenario. Input intents can be generated based on multi-dimensional intent parsing of user input data, and can therefore be complex data structures containing multiple fields. For example, input intents may include fields such as core identifiers, parameters, and attributes.

[0033] In some embodiments, when performing multidimensional intent parsing on the user input data to obtain an intent parsing result including at least one input intent, preprocessing can be performed on the user input data to obtain preprocessed text. This preprocessing includes text cleaning, text segmentation, and context concatenation.

[0034] For example, after obtaining user input data, text preprocessing can be performed first. During text preprocessing, the text content can be cleaned by removing meaningless special characters, emojis, etc. Then, word segmentation and part-of-speech tagging can be performed to obtain a keyword set. Finally, combined with the context, the current input is concatenated with the dialogue history to form a complete contextual text, i.e., the preprocessed text.

[0035] After text preprocessing, the preprocessed text is converted into text vectors, which are dense vectors generated by the language processing model based on the preprocessed text. Then, the classifier in the language processing model is used to calculate the classification probability of the text vectors relative to multiple predefined dimensions of intent, and the input intent is determined based on the classification probability, and intent parsing results are generated based on the input intent.

[0036] For example, an intelligent customer service platform can have a built-in multi-dimensional intent parsing module, which can be used to identify multiple input intents in the user's input text. This multi-dimensional intent parsing module can then incorporate multiple intent parsing paths, such as an end-to-end joint modeling path, a segmentation and recognition cascaded processing path, and a generative understanding path based on a Large Language Model (LLM).

[0037] The end-to-end joint modeling path can employ a pointer network to accurately locate the start and end points of intents and distinguish between primary and non-primary intents in a sentence. Furthermore, prompt-based learning can be utilized to transform multi-intent recognition into a more flexible language generation task.

[0038] The cascaded processing path for segmentation and recognition can obtain the input intent through cascaded processing of multi-intent segmentation and single-intent classification. That is, the multi-dimensional intent parsing module can use a language processing model to segment long sentences containing multiple intents into several independent short sentences or clauses containing only a single intent, and then use a single-intent classification model such as a multi-label classifier to recognize each segmented short sentence.

[0039] LLM-based generative understanding paths can directly process complex user queries through LLM, thereby identifying multiple intents or themes in the utterance at once and automatically generating a unified response or invoking different operations. Furthermore, by designing specific instructions, LLM can perform multiple tasks such as multi-intent recognition and key information extraction while generating answers.

[0040] Therefore, after obtaining the preprocessed text, the multidimensional intent parsing module can perform feature extraction and vectorization on the preprocessed text. By inputting the preprocessed text into a pre-trained language processing model such as BERT, the language processing model converts the preprocessed text into dense vectors that computers can understand, representing the semantic information of the user input data. Then, intent classification is performed, that is, by inputting the generated text vectors into a classifier such as a Softmax layer, the classifier calculates the probability that the text vector belongs to each predefined intent, and at least one intent with a probability score higher than a preset probability threshold is taken as the recognition result.

[0041] As can be seen, the multi-dimensional intent parsing module can extract at least one input intent from user input data, yielding the intent parsing result. When the intent parsing result includes multiple input intents, the current business scenario can be identified as a multi-intent scenario. Similarly, when the intent parsing result includes only one input intent, the current business scenario can be identified as a single-intent scenario.

[0042] S103. Based on the intent parsing results, recall the candidate dialogue set from the question-and-answer knowledge base.

[0043] After obtaining the intent parsing results through multidimensional intent parsing, customer service response content can be matched based on the intent parsing results. At least one response script can be determined in the question-and-answer knowledge base. That is, a set of candidate scripts can be recalled in the question-and-answer knowledge base based on the intent parsing results. The set of candidate scripts includes at least one standard response script that matches the input intent.

[0044] The question-and-answer knowledge base is a structured data repository that stores all business knowledge, common questions, and corresponding solutions for a given business scenario. It can transform fragmented, unstructured business documents into structured, searchable, and standardized data for the system to access during responses. For example, by structuring product manuals, return and exchange policies, and frequently asked questions lists, metadata—knowledge entries—that can be used for retrieval and matching is generated and stored in the question-and-answer knowledge base.

[0045] Standard response scripts are the final output text directly facing users. They can be generated by using a text generation model to transform knowledge entries matched from the question-and-answer knowledge base into natural language in a way that conforms to business specifications and the current conversation context. This ensures that the information is accurate and compliant, and is delivered to users in a human-like, coherent, and polite manner, thereby improving the service experience.

[0046] In some embodiments, when recalling a set of candidate phrases in a question-and-answer knowledge base based on the intent parsing result, a query statement can be generated first based on the intent parsing result, wherein the primary key of the query statement is the input intent; the query statement also includes key entities extracted from the input text.

[0047] Then, following a multi-path recall strategy, standard response phrases matching the input intent are retrieved from the question-and-answer knowledge base based on the query statement. The multi-path recall strategy includes at least one of intent routing retrieval, semantic vector retrieval, and keyword inverted index retrieval.

[0048] The standard response scripts are then merged into the candidate pool, and a candidate script set is selected from the candidate pool. The candidate script set includes at least one standard response script in the candidate pool whose comprehensive relevance score is higher than a preset score threshold.

[0049] For example, when recalling candidate phrase sets, a multi-dimensional query can be constructed first. This involves assembling the structured results of multi-dimensional intent parsing into a single search condition to form the query. When constructing the query, an intent primary key can be set first. This involves using the identified intent tags as precise filtering conditions to determine the knowledge scope within the question-and-answer knowledge base. Then, through slot and entity completion, the extracted key entities are filled into the query for retrieving more detailed processing solutions. To achieve more accurate semantic matching, contextual information can be added to the query. The context-completed query can then serve as a seed for semantic retrieval.

[0050] Then, a multi-path recall strategy is executed based on the query statement. Since the question-answering knowledge base contains tens of thousands of standard question-answer pairs, multiple retrieval paths can be executed in parallel to ensure recall, each pulling candidate phrases from different dimensions. For example, the multi-path recall strategy can include intent-based routing retrieval, semantic vector retrieval, and keyword inverted index retrieval. Intent-based routing retrieval can be a structured path, utilizing intent tags to directly search for standard questions associated with those tags in the question-answering knowledge base index. Semantic vector retrieval can be an unstructured path, transforming the user's question into a vector using deep learning models such as BERT, and then performing an Approximate Nearest Neighbor (ANN) search in the knowledge base to find the top N knowledge entries that are semantically closest. The keyword inverted index retrieval path can extract core keywords from the user input and use tools such as Elasticsearch for traditional BM25 matching to ensure that even when the vector model has biases, relevant phrases can still be accurately matched using keywords.

[0051] After obtaining multiple search results based on the multi-path recall strategy, these results can be merged and coarsely ranked. All candidate phrases recalled through the three paths mentioned above are combined to form a candidate pool. A comprehensive relevance score is then calculated for the search results in the candidate pool. This comprehensive relevance score is calculated as a weighted sum of multiple components: Total Score = α × Intent Matching Weight + β × Vector Similarity + γ × Keyword Hit Rate + δ × Business Priority Weight. Finally, based on the comprehensive relevance scores, the top 5-10 phrases with the highest scores are selected from the candidate pool to form a candidate phrase set.

[0052] like Figure 3 As shown, in some embodiments, after recalling the candidate dialogue set through the question-and-answer knowledge base, the type of the current intent scenario can be determined based on the candidate dialogue set and the intent parsing result. That is, after recalling the candidate dialogue set in the question-and-answer knowledge base based on the intent parsing result, the intent parsing result can be traversed to obtain the number of input intents included in the intent parsing result, and the candidate dialogue set can be traversed to obtain the business action attribute to which the standard reply dialogue in the candidate dialogue set belongs.

[0053] Next, preset interception conditions are obtained, including an intent quantity threshold and a similar business action attribute quantity threshold. The obtained input intent quantity and business action attribute are then compared with the preset interception conditions. When the input intent quantity is less than or equal to the intent quantity threshold, or when the number of standard reply messages belonging to the same business action attribute in the candidate message set is less than or equal to the similar business action attribute quantity threshold, the current intent scenario type can be determined to be a single intent scenario. Therefore, the standard reply message can be directly pushed to the user terminal.

[0054] For example, in the insurance service field, after a user inputs the message "My car was scratched yesterday, how do I report it? Also, when can I get the damage assessment results?", the multi-dimensional intent parsing module can analyze the user input data to obtain the following intent parsing results: Figure 1 Police report guidelines (consult on how to report a police report); Figure 2 For example, inquiring about the progress of damage assessment (consulting on the issuance time of the damage assessment result). The number of input intents in the intent parsing result is 2. Furthermore, the corresponding business action attributes are "reporting process guidance" and "claims progress inquiry," meaning the number of standard response scripts belonging to the same business action attribute in the candidate script set is less than or equal to the threshold for the number of similar business action attributes.

[0055] At this point, independent response scripts are retrieved for each of the two intentions, and corresponding standard response scripts are obtained from the question-and-answer knowledge base. That is, targeting the intentions... Figure 1 The "Reporting Guidelines" provided by the intelligent customer service platform can provide the standard answer: "You can report the incident through the official app, customer service hotline, or WeChat official account." Therefore, the standard script can be directly output: "Hello, after your vehicle was scratched, you can report the incident in the following ways: Log in to the official app and click 'One-Click Reporting,' or call the 24-hour customer service hotline 400-XXX-XXXX, or select 'Claims Service - Reporting' in the WeChat official account menu. Please prepare your policy number and photos of the scene when reporting the incident." Similarly, regarding the meaning Figure 2 When asked about the progress of the damage assessment, the intelligent customer service platform provided the standard answer: "The damage assessment result is usually issued within 1-3 business days after the report is filed." The standard response would then be: "Hello, regarding the damage assessment result, after the surveyor completes the on-site inspection, the result is usually issued within 1-3 business days. You will be notified via SMS at that time, and you can also check the progress in real time on the APP." For example, in the medical field, a user can input text data such as "My child is 3 years old. He started having a fever last night, 38.5 degrees Celsius. He took A to reduce the fever, but it came back this morning. He also has a cough and isn't feeling well." The multi-dimensional intent parsing module can analyze this user input data and obtain the following intent parsing results: primary intent, urgent symptom consultation, such as vomiting, crying, or fear of febrile seizures; secondary intent, coordinating hospital resources, such as requesting early contact with the emergency room; and implicit intent, such as seeking urgent psychological comfort. Similarly, candidate scripts can be retrieved in the same way as in the examples above, such as "High fever accompanied by vomiting and poor mental state indicates an emergency requiring immediate medical attention; please go to the hospital emergency room as soon as possible"; "Generally, fever alone will not directly damage the brain, but persistent high fever accompanied by neurological symptoms requires immediate evaluation by a doctor"; "Online consultation platforms do not provide emergency dispatch services; in case of emergency, please call 120 or go directly to the nearest hospital emergency department"; "On the way to the hospital, you can let the child lie on their side to prevent choking on vomit, and loosen their collar to keep their airway open," etc.

[0056] S104. When the intent parsing result and the candidate speech set meet the preset interception conditions, obtain the rewrite prompt information.

[0057] Based on the intent parsing results and the candidate dialogue set, it is also possible to determine whether the dialogue scenario type corresponding to the current dialogue is a multi-intent scenario through preset interception conditions. Specifically, the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing results and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set. Therefore, when the intent parsing results and the candidate dialogue set meet the preset interception conditions, it is determined that the current scenario is a multi-intent scenario, and thus the push process of the standard reply dialogue can be intercepted, and a rewrite prompt message can be obtained.

[0058] The rewritten prompt information includes structured prompt words generated by encapsulating standard response scripts, user input data, and historical context information of the current dialogue. In other words, the intelligent customer service platform can encapsulate multiple intercepted standard response scripts, user input data, and historical context information of the current dialogue into structured prompt words, which are then input into a large language model for dynamic rewriting.

[0059] like Figure 4As shown, in some embodiments, when the intent parsing result and the candidate dialogue set meet preset interception conditions, when obtaining the rewrite prompt information, the standard reply dialogue can be stopped from being pushed to the user terminal if the number of input intents is greater than the intent number threshold and the number of standard reply dialogues belonging to the same business action attribute in the candidate dialogue set is greater than the same type of business action attribute number threshold, thereby intercepting the standard reply dialogue push process. Then, user input data and the historical context information of the current dialogue are obtained, and the standard reply dialogue, user input data, and historical context information are encapsulated into structured prompt words to obtain the rewrite prompt information.

[0060] For example, given user input data stating "I want to report a crime and request that an investigator be dispatched as soon as possible," after analysis and processing by the multidimensional intent parsing module, the intent parsing result can be determined as: [Intended meaning: "I want to report a crime and request that an investigator be dispatched as soon as possible."] Figure 1 Report the incident (initiate a claim); Figure 2 The system urges an on-site investigation (requesting expedited arrangements). The number of input intents in the intent parsing results is 2, which is greater than the threshold of 1 set in the preset interception conditions. Furthermore, the corresponding business action attributes are all "claims service," meaning the number of standard response scripts belonging to the same business action attribute in the candidate scripts is 2, which is greater than the threshold of 1 for the same type of business action attribute, thus satisfying the preset interception conditions.

[0061] Therefore, standard response scripts can be pre-recalled, specifically the standard answers for "reporting the incident" and "inspection timeframe," respectively. A script fusion interceptor is then triggered to merge the core information from the two standard responses. At this point, the intelligent customer service platform can obtain both standard responses and, combined with user input data and the historical context of the current dialogue, encapsulate them into structured prompts, such as "reporting the incident + urging inspection," which are then input into the large language model for dynamic rewriting.

[0062] S105. Use a large language model to perform dynamic rewriting based on the rewriting prompt information to obtain the rewritten response script.

[0063] After obtaining the rewrite prompt information, the rewrite prompt information can be input into the large language model so that the large language model can perform dynamic rewriting of the response script based on the rewrite prompt information to obtain the rewritten response script.

[0064] like Figure 5 As shown, in some embodiments, when using a large language model to perform dynamic rewriting based on rewriting prompts to obtain rewritten response scripts, the rewriting prompts can be first input into the large language model. Then, the large language model integrates the business logic of the standard response script, as well as captures user emotional characteristics and scenario information based on user input data and historical context information. Finally, dynamic rewriting is performed based on the business logic, user emotional characteristics, and scenario information to obtain the rewritten response scripts.

[0065] For example, multiple intercepted standard response phrases, user input data, and historical context information of the current dialogue can be encapsulated into structured prompts and input into a large language model for dynamic rewriting. The large language model can then integrate the business logic of multiple phrases and perform dynamic rewriting to obtain rewritten response phrases.

[0066] Therefore, the intent parsing result shown in the above embodiments is: Intent Figure 1 Report the incident (initiate a claim); Figure 2 The system urges an on-site investigation (requesting expedited arrangements). Standard response scripts for "reporting process" and "investigation timeframe" can be recalled separately. However, directly splicing or pushing standard response scripts at this point would generate a large amount of repetitive content, such as repeated greetings and identity verifications. Therefore, a script fusion interceptor can be triggered to block the push of standard response scripts.

[0067] Then, the two standard response scripts are combined with user input data and historical context information of the current dialogue to encapsulate them into structured prompts, which are then input into a large language model for dynamic rewriting. The rewritten prompts can include system role and task prompts (e.g., "You are a professional car insurance claims customer service representative, responsible for integrating the following information to generate a response that conforms to business standards and soothes the user's emotions"); historical context prompts, such as a complete record of the previous two rounds of dialogue; current user input prompts, such as "I want to report an accident and request that an investigator be arranged as soon as possible"; and intent parsing result prompts, such as "I want to report an accident and request that an investigator be arranged as soon as possible." Figure 1 "Report a crime"; meaning Figure 2 Standard scripts for "urging an investigation" and pre-recall prompts include: "Hello, after your vehicle has been scratched, you can report the incident in the following ways: log in to the official APP and click 'One-click Report,' or call the 24-hour customer service hotline 400-XXX-XXXX, or select 'Claims Service - Report' in the WeChat official account menu. Please prepare your policy number and photos of the scene when reporting the incident"; and "Hello, an investigator has been dispatched to the scene and is expected to contact you within 30 minutes. Please be patient."

[0068] The above rewriting prompts are combined and input into the large language model, enabling the model to dynamically rewrite the response script. This involves identifying the core information of two standard response scripts, merging them into a concise reply, such as: "Hello, we have expedited the process for an investigator, who is expected to contact you within 30 minutes. You can also check the investigation progress in real time through the app. Please feel free to let me know if you have any other questions." Furthermore, large language models can keenly capture the emotional characteristics and contextual information in the user's original words. For example, if a user is detected to be driving, the rewritten dialogue will automatically include caring phrases such as "be careful." This technology allows the dialogue system to shed its mechanical feel and achieve highly human-like and empathetic dynamic interaction.

[0069] S106. Push the rewritten reply to the user's terminal.

[0070] After dynamically rewriting the response message using a large language model, the rewritten response message can be pushed to the user terminal.

[0071] In some embodiments, when pushing the rewritten response script to the user terminal, a push data stream can also be generated based on the rewritten response script and pushed to the user terminal in a streaming output manner.

[0072] In order to offset the increased time consumption caused by large language model inference, streaming output technology is adopted in engineering implementation, and a lightweight large language model finely tuned for specific business scenarios is used to control the first character response latency of the system to the millisecond level.

[0073] After pushing the rewritten response script to the user's terminal, real business data from the user's terminal can be collected, and the dynamic rewriting strategy of the large language model can be continuously optimized based on the real business data. This real business data includes the user follow-up response rate.

[0074] For example, an intelligent customer service platform can have built-in tracking and monitoring modules and A / B testing modules. Through these modules, real business data such as user follow-up response rates can be collected, forming a data feedback loop, thereby continuously optimizing the rewriting strategy of the large language model.

[0075] By applying the technical solutions of the above embodiments, the multi-intent scenario customer service script processing method described in the above embodiments can adopt a dynamic script fusion and dynamic rewriting scheme, which can reduce transmission redundancy and system maintenance costs. Through semantic deduplication and information dimensionality reduction techniques of the large language model, multiple highly overlapping static scripts are compressed into a single concise reply, effectively reducing redundant data packets in network transmission and lowering the rendering pressure on the client. Furthermore, by utilizing the generalization and understanding capabilities of the large language model's underlying layer, complex hard-coded rules are transformed into extremely simple Prompt instruction configurations, making the system architecture more lightweight, reducing the development costs of later maintenance and iteration, and improving the generalization ability and maintainability of the dialogue system. Moreover, by combining streaming output with a lightweight model, while introducing the large language model inference stage, millisecond-level end-to-end response latency is still maintained, ensuring the high availability and high concurrency processing capabilities of the technical architecture.

[0076] The method also enhances the interactive experience and claims processing efficiency. Through a dialogue fusion interception mechanism and dynamic rewriting of the large language model, it can evolve from rule-based mechanical responses to an empathetic, human-like communication style. Utilizing a natural and fluent interaction, it effectively eliminates user aversion and anxiety caused by automated screen scrolling during the claims process, thus improving customer satisfaction. Simultaneously, precise and empathetic guiding dialogue can more effectively stabilize user emotions, reduce conversation interruptions due to poor communication, and significantly shorten the overall time required for claims document collection, improving the overall operational efficiency of claims processing and laying a solid foundation for future exploration of full-duplex voice claims robots.

[0077] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a multi-intent scenario customer service script processing method. The difference between this method and the above embodiments is that it can improve information transmission efficiency through semantic deduplication and information dimensionality reduction processing. Figure 6 As shown, the method includes: S201. Use a large language model to extract business entities from multiple standard response messages corresponding to the rewritten prompt information; S202. Perform logical merging on multiple standard response scripts to identify the common core information of the multiple standard response scripts; S203. Fill in the common backbone information with the business entities as parallel components to obtain the model transmission information; S204. Generate rewritten response scripts based on the information transmitted by the model.

[0078] When using a large language model to perform dynamic rewriting based on rewriting prompts to obtain rewritten response scripts, the intelligent customer service platform can use the large language model to extract business entities from multiple standard response scripts corresponding to the rewriting prompts, and perform logical merging on multiple standard response scripts to identify the common core information of multiple standard response scripts. Then, the business entities are filled into the common core information as parallel components to obtain the model-transmitted information, and the rewritten response script is generated based on the model-transmitted information.

[0079] For example, intelligent customer service platforms can leverage the natural language understanding and generation capabilities of large language models to perform deep entity extraction and logical merging for repetitive logic such as confirmation phrases and guiding phrases in multiple static customer service messages. Large language models can accurately identify common core information and fill in differentiated business entities as parallel components into the core, such as different document types. This semantic dimensionality reduction process maximizes the conciseness of language and reduces the user's cognitive load while ensuring zero loss of core business information.

[0080] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a multi-intent scenario customer service script processing method. The difference between this method and the above embodiments is that it can construct a strict guardrail mechanism and illusion control strategy to ensure business compliance. Figure 7 As shown, the method includes: S301. Extract core business elements from multiple standard response scripts corresponding to the rewritten prompt message; S302. Generate system prompts based on core business elements; S303. Input the rewrite prompt message and system prompt words into the large language model, and obtain the initial response script output by the large language model; S304. Perform post-validation on the initial response script based on preset validation rules; S305. Output the initial response script that has passed the post-validation as a rewritten response script.

[0081] When using a large language model to perform dynamic rewriting based on rewriting prompts to obtain rewritten response scripts, the intelligent customer service platform can first extract core business elements from multiple standard response scripts corresponding to the rewritten prompts. Then, it generates system prompt words based on these core business elements. These system prompt words are used to define the generation boundaries of the large language model.

[0082] Then, the rewrite prompt message and system prompt words are input into the large language model, along with the initial response script output by the large language model. Based on preset verification rules, a post-verification is performed on the initial response script, whereby the preset verification rules are used to check whether the initial response script contains unauthorized promises. The initial response script that passes the post-verification is then output as the rewritten response script.

[0083] For example, to address the risk of "illusions" generated by large language models during text generation and to adapt to rigorous financial claims scenarios, intelligent customer service platforms can design multiple constraint barriers in their system architecture. Specifically, at the input end, a designed system prompt strictly defines the generation boundaries of the large language model, mandating that it must and can only use the core business elements from the pre-recall script. At the output end, combined with a post-validation module, it ensures that the rewritten response does not contain unauthorized promises, such as claims processing timelines or amounts. Through illusion control strategies and barrier mechanisms, the security pain points of large language models applied to scenarios with high compliance requirements can be effectively addressed.

[0084] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0085] In some embodiments, as a specific implementation of the multi-intent scenario customer service script processing method described in the above embodiments, some embodiments of this application also provide a multi-intent scenario customer service script processing device, such as... Figure 8 As shown, the device includes: The data input module is used to acquire user input data, which is data input based on the user terminal. The multidimensional intent parsing module is used to perform multidimensional intent parsing on the user input data to obtain intent parsing results, the intent parsing results including at least one input intent; The script retrieval module is used to retrieve a set of candidate scripts from the question-and-answer knowledge base based on the intent parsing result. The set of candidate scripts includes at least one standard response script that matches the input intent. An interception decision module is used to obtain rewrite prompt information when the intent parsing result and the candidate dialogue set meet preset interception conditions; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt information includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; The perceptual rewriting module is used to perform dynamic rewriting based on the rewriting prompt information using a large language model to obtain a rewritten response script; The script push module is used to push the rewritten reply script to the user terminal.

[0086] In some embodiments, the multidimensional intent parsing module is specifically used for: The user input data is preprocessed to obtain preprocessed text, including text cleaning, text segmentation, and context concatenation. The preprocessed text is converted into a text vector, which is a dense vector generated by the language processing model based on the preprocessed text. The classifier in the language processing model is used to calculate the classification probability of the text vector relative to multiple predefined dimensional intents; The input intent is determined based on the classification probability, and the intent parsing result is generated based on the input intent.

[0087] In some embodiments, the script recall module is specifically used for: A query statement is generated based on the intent parsing result, wherein the primary key of the query statement is the input intent; the query statement also includes key entities extracted from the input text. According to the multi-path recall strategy, based on the query statement, a standard response script that matches the input intent is retrieved from the question-and-answer knowledge base. The multi-path recall strategy includes at least one of intent routing retrieval, semantic vector retrieval, and keyword inverted index retrieval. The standard response scripts are merged into a candidate pool, and a candidate script set is selected from the candidate pool. The candidate script set includes at least one standard response script in the candidate pool whose overall relevance score is higher than a preset score threshold.

[0088] In some embodiments, the interception decision module is specifically used for: By traversing the intent parsing results, the number of input intents included in the intent parsing results can be obtained; By traversing the candidate script set, the business action attribute to which the standard response script in the candidate script set belongs is obtained; Obtain preset interception conditions, which include a threshold for the number of intents and a threshold for the number of attributes of similar business actions; When the number of input intents is less than or equal to the intent number threshold, or when the number of standard reply scripts belonging to the same business action attribute in the candidate script set is less than or equal to the same type of business action attribute number threshold, the standard reply script is pushed to the user terminal.

[0089] In some embodiments, the interception decision module is specifically used for: When the number of input intents exceeds the intent number threshold, and the number of standard response scripts belonging to the same business action attribute in the candidate script set exceeds the same business action attribute number threshold, the push of the standard response scripts to the user terminal is stopped. Obtain the user input data and the historical context information of the current conversation; The intercepted standard response script, the user input data, and the historical context information are encapsulated into structured prompt words to obtain the rewritten prompt information.

[0090] In some embodiments, the perception rewriting module is specifically used for: Input the rewrite prompt information into the large language model; The business logic of the standard response script is integrated through the large language model, and user emotional characteristics and scenario information are captured based on the user input data and the historical context information. Based on the business logic, the user's emotional characteristics, and the scenario information, dynamic rewriting is performed to obtain the rewritten response script.

[0091] In some embodiments, the perception rewriting module is specifically used for: The large language model is used to extract business entities from the multiple standard response statements corresponding to the rewritten prompt information; Logical merging is performed on multiple standard response statements to identify the common core information of the multiple standard response statements; The business entities are filled into the common backbone information as parallel components to obtain the model transmission information; The rewritten response script is generated based on the information transmitted by the model.

[0092] In some embodiments, the perception rewriting module is specifically used for: Extract core business elements from the multiple standard response scripts corresponding to the rewritten prompt information; System prompts are generated based on the core business elements, and these prompts are used to define the generation boundaries of the large language model. Input the rewrite prompt information and the system prompt words into the large language model, and obtain the initial response script output by the large language model; Based on preset verification rules, post-verification is performed on the initial response script. The preset verification rules are used to verify whether the initial response script contains unauthorized promises. The initial response script that passes the post-validation is output as the rewritten response script.

[0093] In some embodiments, the script delivery module is specifically used for: Generate a push data stream based on the rewritten reply script; The push data stream is pushed to the user terminal in a streaming output manner; Collect real business data fed back from the user terminal, including user follow-up response rate; The dynamic rewriting strategy of the large language model is continuously optimized based on the real business data.

[0094] For specific limitations regarding the device, please refer to the limitations of the method described above, which will not be repeated here. Each module in the device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0095] In some embodiments, a computer device is provided, which may be a server. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of the computer device is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the method on the server side.

[0096] In some embodiments, a computer device is provided, which can be a client. The computer device includes a processor, memory, a network interface, a display screen, and an input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the method on the client side.

[0097] In some embodiments, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire user input data, wherein the user input data is based on data input from the user terminal; Perform multidimensional intent parsing on the user input data to obtain intent parsing results that include at least one input intent; Based on the intent parsing result, a candidate set of dialogues is recalled from the question-and-answer knowledge base, the candidate set of dialogues including at least one standard response dialogue that matches the input intent; When the intent parsing result and the candidate dialogue set meet the preset interception conditions, a rewrite prompt message is obtained; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt message includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script; The rewritten response message is pushed to the user's terminal.

[0098] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire user input data, wherein the user input data is based on data input from the user terminal; Perform multidimensional intent parsing on the user input data to obtain intent parsing results that include at least one input intent; Based on the intent parsing result, a candidate set of dialogues is recalled from the question-and-answer knowledge base, the candidate set of dialogues including at least one standard response dialogue that matches the input intent; When the intent parsing result and the candidate dialogue set meet the preset interception conditions, a rewrite prompt message is obtained; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt message includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script; The rewritten response message is pushed to the user's terminal.

[0099] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0101] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0102] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0103] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for processing customer service scripts in multi-intent scenarios, characterized in that, The method includes: Acquire user input data, wherein the user input data is based on data input from the user terminal; Perform multidimensional intent parsing on the user input data to obtain intent parsing results that include at least one input intent; Based on the intent parsing result, a candidate set of dialogues is recalled from the question-and-answer knowledge base, the candidate set of dialogues including at least one standard response dialogue that matches the input intent; When the intent parsing result and the candidate dialogue set meet the preset interception conditions, a rewrite prompt message is obtained; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt message includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script; The rewritten response message is pushed to the user's terminal.

2. The method according to claim 1, characterized in that, After recalling the candidate phrase set in the question-answering knowledge base based on the intent parsing results, the method further includes: Iterate through the intent parsing results to obtain the number of input intents included in the intent parsing results; Traverse the candidate script set to obtain the business action attribute to which the standard response script in the candidate script set belongs; Obtain preset interception conditions, which include a threshold for the number of intents and a threshold for the number of attributes of similar business actions; When the number of input intents is less than or equal to the intent number threshold, or when the number of standard reply scripts belonging to the same business action attribute in the candidate script set is less than or equal to the same type of business action attribute number threshold, the standard reply script is pushed to the user terminal.

3. The method according to claim 2, characterized in that, When the intent parsing result and the candidate speech set meet the preset interception conditions, a rewrite prompt message is obtained, including: When the number of input intents exceeds the intent number threshold, and the number of standard response scripts belonging to the same business action attribute in the candidate script set exceeds the same business action attribute number threshold, the push of the standard response scripts to the user terminal is stopped. Obtain the user input data and the historical context information of the current conversation; The standard response script, the user input data, and the historical context information are encapsulated into structured prompt words to obtain the rewrite prompt information.

4. The method according to claim 1, characterized in that, Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script, including: Input the rewrite prompt information into the large language model; The business logic of the standard response script is integrated through the large language model, and user emotional characteristics and scenario information are captured based on the user input data and the historical context information. Based on the business logic, the user's emotional characteristics, and the scenario information, dynamic rewriting is performed to obtain the rewritten response script.

5. The method according to claim 1, characterized in that, Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script, including: The large language model is used to extract business entities from the multiple standard response statements corresponding to the rewritten prompt information; Logical merging is performed on multiple standard response statements to identify the common core information of the multiple standard response statements; The business entities are filled into the common backbone information as parallel components to obtain the model transmission information; The rewritten response script is generated based on the information transmitted by the model.

6. The method according to claim 1, characterized in that, Dynamic rewriting is performed using a large language model based on the rewriting prompt information to obtain a rewritten response script, including: Extract core business elements from the multiple standard response scripts corresponding to the rewritten prompt information; System prompts are generated based on the core business elements, and these prompts are used to define the generation boundaries of the large language model. Input the rewrite prompt information and the system prompt words into the large language model, and obtain the initial response script output by the large language model; Based on preset verification rules, post-verification is performed on the initial response script. The preset verification rules are used to verify whether the initial response script contains unauthorized promises. The initial response script that passes the post-validation is output as the rewritten response script.

7. The method according to claim 1, characterized in that, Pushing the rewritten response message to the user terminal includes: Generate a push data stream based on the rewritten reply script; The push data stream is pushed to the user terminal in a streaming output manner; Collect real business data fed back from the user terminal, including user follow-up response rate; The dynamic rewriting strategy of the large language model is continuously optimized based on the real business data.

8. A multi-intent scenario customer service script processing device, characterized in that, The device includes: The data input module is used to acquire user input data, which is data input based on the user terminal. The multidimensional intent parsing module is used to perform multidimensional intent parsing on the user input data to obtain intent parsing results, the intent parsing results including at least one input intent; The script retrieval module is used to retrieve a set of candidate scripts from the question-and-answer knowledge base based on the intent parsing result. The set of candidate scripts includes at least one standard response script that matches the input intent. An interception decision module is used to obtain rewrite prompt information when the intent parsing result and the candidate dialogue set meet preset interception conditions; the preset interception conditions are used to intercept the push process of the standard reply dialogue based on the number of input intents in the intent parsing result and the business action attribute to which the standard reply dialogue belongs in the candidate dialogue set; the rewrite prompt information includes structured prompt words generated by encapsulating the standard reply dialogue, the user input data, and the historical context information of the current dialogue; The perceptual rewriting module is used to perform dynamic rewriting based on the rewriting prompt information using a large language model to obtain a rewritten response script; The script push module is used to push the rewritten reply script to the user terminal.

9. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images