Method, device, electronic device and processor for processing service text of vehicle

Abstract

This application discloses a method, apparatus, electronic device, and processor for processing service text for vehicles. The service text represents the process of a first object providing services for a vehicle to a second object. The service text includes multiple text segments. The method includes: in response to obtaining the service text, determining the service type of each text segment, wherein different service types correspond to different service contents provided by the first object to the second object; invoking a detection strategy corresponding to the service type to detect the text segments, obtaining a first detection sub-result for each text segment; in response to the first detection sub-result indicating that the accuracy of the performance index is less than an accuracy threshold, detecting the text segments again, obtaining a second detection sub-result for each text segment; and determining the detection result of the service text based on the multiple second detection sub-results corresponding to the multiple text segments. This application solves the technical problem of low accuracy in processing service text for vehicles.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and natural language processing technology, and more specifically, provides a method, apparatus, electronic device, and processor for processing service text for vehicles. Background Technology

[0002] Currently, in the calculation of vehicle sales rebates, the compliance of service texts (sales dialogues) must be checked to ensure that the Standard Operating Procedure (SOP) is followed.

[0003] In related technologies, manual keyword matching can be used for service text, but this has a high false positive rate and fails to identify intent. Third-party cloud application programming interfaces (APIs) are costly, logically black-box, and difficult to customize. Large, standalone models are prone to creating illusions of active and passive behavior when processing long texts, and they consume a lot of tokens. Therefore, the technical problem of low accuracy in processing vehicle service texts still exists.

[0004] There is currently no effective solution to the aforementioned technical problems. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, and processor for processing service text for vehicles, to at least solve the technical problem of low accuracy in processing service text for vehicles.

[0006] According to one aspect of the embodiments of this application, a method for processing service text of a vehicle is provided. The service text represents the process of a first object providing services for a vehicle to a second object. The service text includes multiple text fragments. The method includes: in response to obtaining the service text, determining the service type of each text fragment, wherein different service types correspond to different service contents provided by the first object to the second object; invoking a detection strategy corresponding to the service type to detect the text fragments and obtain a first detection sub-result of the text fragments, wherein the detection strategy represents the rules for detecting the standardized performance indicators of the service content in the text fragments; in response to the first detection sub-result indicating that the accuracy of the standardized performance indicator is less than an accuracy threshold, detecting the text fragments and obtaining a second detection sub-result of the text fragments, wherein the second detection sub-result indicates that the accuracy of the standardized performance indicator is greater than or equal to the accuracy threshold; and determining a detection result of the service text based on the multiple second detection sub-results corresponding to the multiple text fragments, wherein the detection result represents the standardized performance indicator of the service provided by the first object to the second object.

[0007] Optionally, in response to obtaining the service text, determining the service type of the text fragment includes: in response to obtaining the service text, segmenting the service text based on the dialogue rounds between the first object and the second object to obtain multiple text fragments; and mapping the multiple text fragments to the corresponding service types.

[0008] Optionally, mapping multiple text fragments to corresponding service types includes: invoking a classification agent to determine the service stage to which the multiple text fragments belong; in response to multiple service stages corresponding to consecutive multiple text fragments being in an abnormal state, using the classification agent to adjust at least one of the consecutive multiple service stages to obtain an adjusted service stage, wherein the adjusted consecutive multiple service stages are in a normal state; and mapping the adjusted service stage to a service type.

[0009] Optionally, the method further includes: acquiring dialogue text characters in a text fragment; and determining a detection strategy based on the number and service type of the dialogue text characters.

[0010] Optionally, the detection strategy includes a first detection strategy and a second detection strategy. The first detection strategy represents the rules for detecting individual characters in the dialogue text, and the second detection strategy represents the rules for detecting the entire text segment. The detection strategy corresponding to the service type is invoked to detect the text segment, resulting in a first detection sub-result. This includes: responding to a situation where the number of dialogue text characters is greater than or equal to a threshold, invoking the first detection strategy of the quality inspection agent to perform switching and vectorization processing on the dialogue text characters, obtaining a processing result; invoking the first detection result and determining the first detection sub-result based on the processing result; and responding to a situation where the number of dialogue text characters is less than the threshold, invoking the second detection strategy of the quality inspection agent to directly detect the text segment, obtaining the first detection sub-result.

[0011] Optionally, the method further includes: invoking a review agent to review the first detection sub-result and obtain a review result, wherein the review result is used to represent the detection quality of the first detection sub-result; and in response to the detection quality being lower than a quality threshold, determining that the accuracy of the first detection sub-result is less than an accuracy threshold.

[0012] Optionally, in response to the first detection sub-result indicating that the accuracy of the performance index of the detection specification is less than the accuracy threshold, the text segment is detected to obtain a second detection sub-result of the text segment, including: in response to the first detection sub-result indicating that the accuracy of the performance index of the detection specification is less than the accuracy threshold, the review agent is invoked to review the text segment to obtain the second detection sub-result.

[0013] Optionally, the method further includes: invoking an insight agent to generate a management policy for a first object based on the detection results, wherein the management policy is used to represent rules for managing the first object to improve the standard performance metrics of the first object in providing services to the second object.

[0014] According to another aspect of the embodiments of this application, a processing apparatus for service text of a vehicle is also provided. The apparatus may include: a first determining unit, configured to determine the service type of a text segment in response to obtaining service text, wherein different service types correspond to different service contents of a service provided by a first object to a second object; a first detecting unit, configured to invoke a detection strategy corresponding to the service type to detect the text segment and obtain a first detection sub-result of the text segment, wherein the detection strategy is used to represent the rules for detecting the standardized performance indicators of the service content in the text segment; a second detecting unit, configured to detect the text segment in response to the first detection sub-result indicating that the accuracy of the standardized performance indicator is less than an accuracy threshold, and obtain a second detection sub-result of the text segment, wherein the second detection sub-result indicates that the accuracy of the standardized performance indicator is greater than or equal to the accuracy threshold; and a second determining unit, configured to determine the detection result of the service text based on multiple second detection sub-results corresponding to multiple text segments, wherein the detection result is used to represent the standardized performance indicators of the service provided by the first object to the second object.

[0015] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided. The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the methods described in the embodiments of this application.

[0016] According to another aspect of the embodiments of this application, a processor is also provided. This processor is used to run a program, wherein the program executes the methods described in the embodiments of this application during runtime.

[0017] According to another aspect of the embodiments of this application, an electronic device is also provided. The electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the methods described in the embodiments of this application.

[0018] According to another aspect of the embodiments of this application, a computer program product is also provided. This computer program product includes a computer program that, when executed by a processor, implements the methods described in the embodiments of this application.

[0019] In this embodiment, if it is necessary to process the service text of a vehicle, the service type corresponding to each text segment in the service text can be determined. The detection strategy corresponding to the service type is invoked to detect the standardization indicators of the service content in the text segment, obtaining a first detection sub-result. If the accuracy of the first detection sub-result in detecting the standardization indicator is less than the accuracy threshold, the text segment can be detected to obtain a second detection sub-result, thereby improving the accuracy of the standardization performance indicators. Based on the second detection sub-results of each text segment, the detection result of the entire service text can be determined. In other words, in this embodiment, by dividing the service text into multiple text segments and dynamically invoking the corresponding detection strategy according to the service type of each text segment, fine-grained, scenario-based compliance detection is achieved, effectively avoiding misjudgment of intent and token waste caused by excessively long context in the single model. When the accuracy of the first detection sub-result is lower than the accuracy threshold, a secondary high-precision detection mechanism is automatically triggered to improve the accuracy of the judgment of the standardization performance indicators; finally, the second detection sub-results of each text segment are aggregated to form the overall service standardization detection result. This embodiment significantly improves the accuracy of service text quality inspection through segmented processing, service type adaptation, and dual verification. It avoids the high misjudgment rate caused by coarse-grained analysis, black-box reasoning, and long text loss in related technologies, solves the technical problem of low accuracy in processing vehicle service text, and achieves the technical effect of improving the accuracy of vehicle service text processing. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 This is a schematic diagram of a vehicle service text processing application scenario according to an embodiment of this application;

[0022] Figure 2 This is a flowchart of a method for processing service text for a vehicle according to an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of an offline dialogue quality inspection system based on a multi-agent collaboration and arbitration mechanism according to an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of a vehicle service text processing apparatus according to an embodiment of this application;

[0025] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application;

[0026] Figure 6This is a schematic diagram of a computing device according to an embodiment of this application. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] Figure 1 This is a schematic diagram illustrating a vehicle service text processing application scenario according to an embodiment of this application, such as... Figure 1 As shown, the above scenario may include terminal device 10, network 20, and server 30. Terminal device 10 can acquire the voice dialogue between a first and second party based on user input. This voice dialogue can be sent to vehicle 30 via network 20. Vehicle 30 can also convert the voice dialogue into service text. At this point, server 30 needs to execute steps S102 to S108 to implement the service text processing for the vehicle.

[0030] The following steps can be performed by vehicle 30: Step S102, in response to obtaining service text, determine the service type of the text fragment; Step S104, call the detection strategy corresponding to the service type to detect the text fragment and obtain the first detection sub-result of the text fragment; Step S106, in response to the first detection sub-result indicating that the accuracy of the performance index is less than the accuracy threshold, detect the text fragment and obtain the second detection sub-result of the text fragment; Step S108, based on the multiple second detection sub-results corresponding to multiple text fragments, determine the detection result of the service text.

[0031] In this embodiment, through steps S102 to S108, the service text is segmented into multiple text fragments. Corresponding detection strategies are dynamically invoked based on the service type of each text fragment, achieving fine-grained, scenario-based compliance detection. This effectively avoids misjudgment of intent and token waste caused by excessively long context in individual models. When the accuracy of the first detection sub-result is lower than the accuracy threshold, a secondary high-precision detection mechanism is automatically triggered to improve the accuracy of judging compliance performance indicators. Finally, the second detection sub-results of each text fragment are aggregated to form the overall compliance detection result for the service. This embodiment significantly improves the accuracy of service text quality inspection through segmented processing, service type adaptation, and dual verification. It avoids the high misjudgment rate caused by coarse-grained analysis, black-box reasoning, and long text errors in related technologies, solving the technical problem of low accuracy in processing vehicle service text and achieving the technical effect of improving the accuracy of vehicle service text processing.

[0032] According to an embodiment of this application, an embodiment of a method for processing service text of a vehicle is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] Figure 2 This is a flowchart of a method for processing service text for a vehicle according to an embodiment of this application, such as... Figure 2 As shown, the service text is used to represent the process by which a first object provides a service for a vehicle to a second object. The service text includes multiple text fragments, and the method may include the following steps:

[0034] Step S202: In response to obtaining the service text, determine the service type of the text fragment.

[0035] In the technical solution provided by step S202 of this application, different service types correspond to different service contents provided by the first object to the second object.

[0036] Optionally, the service text can be a structured text representing the entire process of a voice dialogue between a first party (e.g., a sales consultant) and a second party (e.g., a customer) providing vehicle-related sales services, converted through speech recognition. This service text can include a complete transcript of the sales dialogue, covering the entire interactive chain from customer reception at the store, needs assessment, product explanation, objection handling, test drive invitation to closing the deal. For example, the service text can be a long text (e.g., over 3000 words).

[0037] Optionally, the first party can refer to the party that proactively provides services during the vehicle service (e.g., sales service) process. For example, it could be a sales consultant or service personnel at a car dealership. The actions of the first party must comply with the company's SOPs. All wording, guidance, promises, and service actions in the service document are issued by the first party.

[0038] Optionally, the second party can refer to the customer receiving the service, that is, a potential car buyer or a consumer who has entered the sales process. The aforementioned second party raises needs, expresses doubts, and provides feedback in the service document, and is the receiving end of the service response. The words and actions of the aforementioned second party can serve as a reference for judging whether the service provided by the first party is "proactive" and "compliant".

[0039] Optionally, the text fragments in the service text can be the smallest semantic units formed by logically segmenting the complete service text into single-turn dialogues (i.e., an interaction unit consisting of one party speaking and the other party responding). Each text fragment can contain one sentence or several consecutive dialogues, corresponding to an independent information exchange behavior. It is the basic processing unit for intent recognition, journey classification, and accurate detection. The boundaries of the above text fragments are clear, which facilitates parallel analysis and context tracking.

[0040] Optionally, service type can refer to the business journey category identified by a classification agent through semantic understanding and business context analysis of various text fragments, reflecting the stage of the sales service process. For example, service types may include, but are not limited to: in-store pickup, needs assessment, product explanation, objection handling, and test drive invitation. Different service types correspond to different SOP rule sets, which are the prerequisites for triggering differentiated detection strategies, ensuring that quality inspection behavior accurately matches the business scenario and avoiding misjudgments caused by mismatched rules. For example, the aforementioned service type can be a business journey category.

[0041] In this embodiment, if it is necessary to process the service text of a vehicle, the service type of each text fragment in the service text can be determined.

[0042] Optionally, in response to the acquisition of service text, the service text can be preprocessed, segmenting it according to dialogue rounds to form multiple text fragments. Each text fragment corresponds to a complete or continuous statement issued by the first or second party, with the speaker switching as the boundary. For example, a salesperson speaking followed by a customer response, and then the salesperson speaking again, constitute two independent text fragments. The segmentation criteria may include the speaker identifiers inherent in the speech-to-text, punctuation and sentence breaks, and semantic integrity, ensuring that each fragment is semantically independent and facilitates independent subsequent analysis.

[0043] Optionally, after obtaining the set of text fragments, a pre-trained lightweight classification agent is invoked. This agent can be fine-tuned based on the business corpus and is capable of recognizing typical service journeys in car sales. The entire content of each text fragment is input into the classification agent, and through semantic encoding and intent matching, the business journey category to which the text fragment belongs, i.e., the service type, is output. This classification process does not rely on keyword matching but is based on the contextual semantic understanding of the text fragments. For example, identifying "Do you usually drive mainly in the city?" as a demand discovery, "This car has more space than the XXX model" as objection handling, and "Would you like to take a test drive now?" as a test drive invitation.

[0044] Optionally, to eliminate accidental fluctuations during the classification process, a sequence smoothing algorithm can be introduced after classification to perform temporal consistency checks on the service types of multiple consecutive text segments. If a text segment is classified as "greeting upon arrival," but the text segments before and after it are classified as "product explanation" with extremely short time intervals, it is judged as a noise misclassification and automatically corrected to "product explanation" consistent with the preceding and following text segments (context), ensuring the continuity and logical rationality of the service.

[0045] In this embodiment, fine-grained analysis is achieved by segmenting dialogue turns to avoid contextual confusion caused by blind full-text detection; semantic classification agents are used to replace keyword rules, significantly improving the recognition accuracy of ambiguous expressions, spoken variations, and complex dialogues; and temporal smoothing algorithms are used to suppress model classification jitter, ensuring the coherence and authenticity of the business journey sequence.

[0046] Step S204: Invoke the detection strategy corresponding to the service type to detect the text fragment and obtain the first detection sub-result of the text fragment.

[0047] In the technical solution provided in step S204 of this application, the detection strategy can be used to represent the rules for detecting the standard performance indicators of service content in text fragments.

[0048] Optionally, the detection strategy can refer to a set of compliance detection rules and execution logic customized for a specific service type. This can guide the quality inspection agent to judge whether the service behaviors presented in the corresponding text fragments comply with the enterprise's SOP standards. Each service type corresponds to an independent detection strategy. These detection strategies may include: triggered keyword semantic patterns, prohibited behavioral expressions, judgment conditions for active / passive voice, and context-dependent compliance elements. For example, the detection strategy in the "demand mining" stage can require sales staff to proactively inquire about the customer's car purchase budget, usage scenarios, or replacement intentions; failure to do so is considered a failure to meet the standard. In the "objection handling" stage, the detection strategy can require sales staff to provide targeted responses to customer comparisons of competing products, rather than avoiding the issue or giving vague statements. The detection strategy is not a fixed regular expression, but rather a semantic rule encapsulated through structured prompts. It is executed by a lightweight, large model as the quality inspection agent, supporting semantic understanding, intent recognition, and logical reasoning, thereby achieving accurate compliance judgment of complex sales scripts.

[0049] Optionally, the first detection sub-result can refer to the preliminary conclusion output by the quality inspection agent after executing the detection strategy. The aforementioned first detection sub-result can include three elements: whether the rule is matched, the corresponding original text quotation, and a brief judgment explanation. The aforementioned first detection sub-result can be used to reflect the judgment made by the agent based on rules and semantic understanding under the current context. Essentially, it is a preliminary inspection opinion / result, which may be biased due to model illusion, lack of context, or misjudgment of active / passive relationships. For example, when the rule requires "the salesperson proactively provides drinks," but the original text of the fragment is "Customer: Can you give me a glass of water? Salesperson: Sure, coming right away," the preliminary inspection result may be misjudged as a match because only the action of "providing" is recognized, while the core constraint of "proactively" is not. The first detection sub-result is not a final conclusion, but rather serves as input for subsequent arbitration. Its value lies in providing traceable and auditable reasoning traces, preserving the original judgment path for subsequent review and correction.

[0050] Optionally, performance metrics can refer to quantitative or qualitative evaluation dimensions used in specific service scenarios to measure whether a service provider (such as a car sales consultant) performs service actions in accordance with preset standard procedures, codes of conduct, or business compliance requirements. For example, the aforementioned performance metrics can be used to indicate whether a sales conversation conforms to preset business rules and SOP (Standard Operating Procedure) script standards.

[0051] In this embodiment, after determining the service type of each text fragment in the service text, the detection strategy corresponding to the service type can be invoked to detect the text fragment and obtain the first detection sub-result of the text fragment.

[0052] Optionally, during the process of calling the detection strategy corresponding to the service type to detect the text fragment and obtain the first detection sub-result, a set of detection rules that strictly match the service type is dynamically loaded from a pre-set strategy library according to the service type to which the text fragment belongs. Each detection rule is encapsulated in the form of a structured Prompt, including elements such as rule title, semantic definition, positive compliance standards, negative violation standards, whether active behavior is required, and whether it depends on context, and it is mandatory for the model to output a traceable thought process and original text citations, prohibiting the direct presentation of conclusions.

[0053] Optionally, the text fragment, along with the loaded detection strategy Prompt, is sent to the quality inspection agent as input. Upon receiving the instruction, the quality inspection agent first parses the semantic constraints in the rules. For example, it determines whether "proactively inquiring" is valid, excluding cases where the customer initiates the request; when identifying "mentioning specific competitor names," it distinguishes between salesperson-initiated comparisons and customer-initiated mentions; and it confirms whether the "test drive invitation" includes a clear time, location, or guiding message. Throughout this process, the model does not rely on keyword matching but rather on semantic understanding, intent inference, and logical relationship analysis, combined with the dialogue context, to make a comprehensive judgment.

[0054] Optionally, to avoid illusions and misjudgments, the detection strategy can require the model to output a chain of thought, i.e., a step-by-step reasoning process. For example: "The customer first says 'I'm a little thirsty,' which is an expression of need; the salesperson responds 'I'll get you a glass of water,' which is a passive response; the rule requires 'actively asking,' therefore the condition is not met." Simultaneously, the model can use matched quotes from the original text that directly support the judgment, ensuring that each first detection sub-result is supported by the original text corpus and eliminating unfounded speculation.

[0055] In this embodiment, the above method achieves scenario-based and differentiated execution of quality inspection tasks by precisely binding service types and detection strategies, avoiding misjudgments caused by the one-size-fits-all approach of traditional general models; by forcing the model to output interpretable reasoning chains and original text evidence through structured prompts, the illusion rate is significantly reduced and the credibility of the results is enhanced; by using parallel processing and conditional retrieval-Augmented Generation (RAG) mechanisms, both processing efficiency and the completeness of long text understanding are taken into account, and the computational cost is significantly reduced without sacrificing accuracy; the first detection sub-result generated at the end has the characteristics of being traceable, reviewable, and reproducible, providing a high-quality, low-noise original judgment basis for subsequent multi-layer arbitration mechanisms, which is the core supporting link for building a high-accuracy automated quality inspection system.

[0056] Step S206: In response to the fact that the accuracy of the performance index of the first detection sub-result is less than the accuracy threshold, the text segment is detected to obtain the second detection sub-result of the text segment.

[0057] In the technical solution of step S206 of this application, the second detection sub-result indicates that the accuracy of the performance index is greater than or equal to the accuracy threshold.

[0058] Optionally, the second detection sub-result can refer to a compliance judgment result with higher credibility and business accuracy, formed after multiple rounds of intelligent agent collaborative review and arbitration mechanisms, provided that the first detection sub-result determines that the accuracy is below a preset threshold, has potential misjudgments, or has logical flaws. This second detection sub-result is not a simple re-examination of the original text, but rather an authoritative conclusion generated based on adversarial reasoning and multi-role verification processes, and has undergone logical verification and factual review.

[0059] In this embodiment, after detecting the text segment and obtaining the first detection sub-result, if the accuracy of the performance index indicated by the first detection sub-result is less than the accuracy threshold, the text segment can be detected again for greater precision to obtain the second detection sub-result.

[0060] Optionally, during the accuracy evaluation of the first detection sub-result, the first detection sub-result output by the quality inspection agent can be received. The first detection sub-result is automatically scored according to preset accuracy evaluation rules. The evaluation criteria may include: whether there are logical gaps in the reasoning process, such as misjudging "customer proactively mentioned" as "sales proactively inquired"; whether the cited MatchedQuote truly supports the conclusion, and whether there is any misinterpretation or contextual exclusion; and whether the remarks contain a clear interpretation of the core verbs of the rule (e.g., "proactively," "must," "prohibited"). If the built-in evaluation logic determines that the result has at least one high-risk misjudgment feature, the first detection sub-result is marked as low confidence, that is, the accuracy is below the accuracy threshold.

[0061] Optionally, when the accuracy of the first detection sub-result is determined to be lower than the accuracy threshold, the review agent is automatically activated. The review agent's Prompt can be set to an adversarial role, with the core instruction being "You are a rigorous quality inspector who must find any logical loopholes, semantic substitutions, or factual deviations in the first detection result, especially hidden errors such as active / passive confusion, misaligned intent, and missing context." The review agent receives three inputs: the full dialogue text of the original service text, the first detection sub-result, and the original rule definition. The review agent does not rely on the model's initial judgment of the first detection sub-result but independently re-analyzes the semantics, focusing on verifying whether the keywords in the rule are correctly executed. For example, if the rule requires "the salesperson to proactively invite a test drive," but the initial detection result cites "the customer asked if a test drive could be arranged," the review agent will identify that the core constraint "proactively" is not satisfied.

[0062] Optionally, the review agent outputs structured opinions (review opinions). The second detection sub-result mentioned above may include a Boolean field (has_problem) and a detailed description of the error (problem_description). If an error is found, has_problem is true, and the "error type," "error basis," and "correct judgment method" are clearly indicated. For example, "Problem: Misjudged that the salesperson proactively offered a drink. Basis: The customer's original words were 'Could you give me a glass of water?', and the salesperson responded 'I'll get it,' which is a passive response. The rule requires 'proactive inquiry,' meaning the salesperson must first ask the customer if they need anything, which is not reflected here. The correct judgment should be Matched: false." This opinion serves as the sole input basis for arbitration.

[0063] Optionally, when the review comments point out a problem, the original dialogue, the first detection sub-result, and the review comments are automatically input into the arbitration agent. The arbitration agent's prompt is: "You are a final arbitrator, judging which side's viewpoint is more consistent with the facts based solely on the original dialogue content and rule definitions. You must ignore the model's subjective bias and only look at the evidence." The arbitration agent does not perform new reasoning but focuses on "whether the original corpus supports the rule requirements." For example, if the rule requires "actively asking questions," then the dialogue must contain statements where the seller actively raises relevant questions; otherwise, it is non-compliant. The arbitration agent does not rely on any external knowledge and takes the original text as the sole source of truth. The review agent generates the second detection sub-result. The arbitration agent outputs the final JavaScript Object Notation (JSON) result, i.e., the second detection sub-result. The second detection sub-result may include Matched, the only supporting statement confirmed by arbitration (MatchedQuote), and the arbitration conclusion statement (Remarks). Based on the reasoning completeness, original text citation accuracy, and semantic consistency of the second detection sub-result, the accuracy of the second detection sub-result is automatically evaluated, and it is confirmed that the above accuracy is greater than or equal to the preset threshold, thereby meeting the system's requirement for high-confidence output.

[0064] In this embodiment, the above method achieves automatic interception of low-quality judgments through an accuracy threshold triggering mechanism, avoiding the pollution of downstream statistics by erroneous results and significantly improving the reliability of the overall quality inspection results. Through a review-arbitration two-layer error correction architecture, the "review thinking" of human quality inspectors is transformed into a repeatable and scalable artificial intelligence (AI) process, which can correct high-difficulty logical errors without human intervention, reducing the illusion rate to an industrially usable level. A mechanism for automatic upgrade of low confidence is constructed. Each arbitration correction provides high-quality error correction samples for future quality inspection agents, promoting the model to continuously improve its inference robustness without retraining.

[0065] Step S208: Based on multiple second detection sub-results corresponding to multiple text fragments, determine the detection result of the service text.

[0066] In the technical solution of step S208 of this application, the test result can be used to represent the standard performance index of the service provided by the first object to the second object.

[0067] Optionally, the detection result may refer to the global, structured, and interpretable final judgment on the performance of the normative indicators in the complete service text, formed by systematically aggregating and semantically integrating multiple second detection sub-results corresponding to multiple text fragments.

[0068] In this embodiment, after detecting the text fragments and obtaining the second detection sub-results of the text fragments, the detection result of the service text can be determined based on the multiple second detection sub-results corresponding to multiple text fragments.

[0069] Optionally, multiple second detection sub-results are received from the arbitration module, each output in a uniform JSON format. By performing format validation and integrity checks on the above second detection sub-results, entries with missing key fields or abnormal formats are removed, ensuring that the second detection sub-results are high-quality data that has been corrected by arbitration, meets the confidence level, and is traceable.

[0070] Optionally, based on the JourneyStage field carried in each second detection sub-result, the scattered second detection sub-results are rearranged and aligned according to the time sequence of the business journey to restore the complete timeline of the service dialogue. For example, the second detection sub-results of stages such as "greeting customers upon arrival," "demand mining," "product explanation," "objection handling," and "closing the sale" are sorted according to the order in which the dialogue occurs to form a logically coherent service journey chain.

[0071] Optionally, based on the semantic attributes of the standardized indicators, the results of each second detection sub-test can be categorized and aggregated. For example, indicators related to "proactiveness" (such as proactively inquiring about the budget, proactively inviting test drives, and proactively offering drinks) can be categorized into one group; indicators related to "information completeness" (such as whether the three-guarantee policy is introduced, whether financial solutions are explained, etc.) can be categorized into another group; and indicators related to "compliance" (such as whether unofficial benefits are promised, whether competitors are disparaged, etc.) can be categorized into a third group.

[0072] Optionally, statistical aggregation is performed on each cluster of performance metrics. For each metric cluster, the corresponding achievement rate is calculated, which is the number of sub-results with a match of true within the metric cluster divided by the total number of sub-results in that cluster, forming a quantitative score. Simultaneously, abnormal patterns are identified: if a certain metric fails to achieve its target in multiple consecutive journey stages (e.g., no proactive inquiry about the budget was made in either the "requirement mining" or "product explanation" stages), it is marked as "systematic deficiency"; if a certain metric repeatedly contradicts itself in multiple text fragments (e.g., once judged as "proactive," another as "passive," but after arbitration both are "mismatched"), it is marked as "unstable behavior."

[0073] Optionally, semantic association analysis can be introduced during the aggregation process. For example, if "not proactively inquiring about the budget" is repeatedly flagged as a miss in the "demand mining" phase, while customer feedback about "the price being too high" frequently appears in the "objection handling" phase, a causal relationship will be automatically established: budget not determined → customer later raises price objections. This relationship is not a statistical correlation, but rather a reasoning binding based on manually generated semantic logic in Remarks (e.g., "due to not confirming the budget in advance, the customer is price-sensitive"), connecting isolated detection points into a business logic chain.

[0074] Optionally, all the above aggregate analysis results can be integrated into a single structured final detection result.

[0075] In this embodiment, the above method achieves a leap from fragment-level judgment to service-level evaluation, upgrading the detection results from scattered compliance checks to a quality profile of the entire service process, thus enhancing the systematic nature of management decisions. Through journey structure reconstruction and semantic aggregation, the detection results are ensured to be consistent with the actual business process, avoiding analytical distortion caused by text fragmentation. The introduction of abnormal pattern recognition and causal reasoning gives the detection results diagnostic capabilities, not only pointing out "where the error occurred," but also explaining "why the error occurred" and "what the possible consequences are," greatly enhancing the interpretability and action orientation of the results. The detection results are output in a dual form: structured data and natural language suggestions, satisfying both the system's automated processing needs and directly serving human management decisions, achieving a seamless integration of technical capabilities and business value.

[0076] In steps S202 to S208 of this application, if it is necessary to process the service text of the vehicle, the service type corresponding to each text segment in the service text can be determined. The detection strategy corresponding to the service type is invoked to detect the standardization indicators of the service content in the text segment, obtaining a first detection sub-result. If the accuracy of the first detection sub-result in detecting the standardization indicator is less than the accuracy threshold, the text segment can be detected to obtain a second detection sub-result, thereby improving the accuracy of the standardization performance indicators. Based on the second detection sub-results of each text segment, the detection result of the entire service text can be determined. In other words, in this embodiment, by dividing the service text into multiple text segments and dynamically invoking the corresponding detection strategy according to the service type of each text segment, fine-grained, scenario-based compliance detection is achieved, effectively avoiding misjudgment of intent and token waste caused by excessively long context in the single model. When the accuracy of the first detection sub-result is lower than the accuracy threshold, a secondary high-precision detection mechanism is automatically triggered to improve the accuracy of the judgment of the standardization performance indicators; finally, the second detection sub-results of each text segment are aggregated to form the overall service standardization detection result. This embodiment significantly improves the accuracy of service text quality inspection through segmented processing, service type adaptation, and dual verification. It avoids the high misjudgment rate caused by coarse-grained analysis, black-box reasoning, and long text loss in related technologies, solves the technical problem of low accuracy in processing vehicle service text, and achieves the technical effect of improving the accuracy of vehicle service text processing.

[0077] The method described in this embodiment will be further described below.

[0078] As an optional embodiment, step S202, in response to obtaining the service text, determines the service type of the text fragment, including: in response to obtaining the service text, segmenting the service text based on the dialogue rounds between the first object and the second object to obtain multiple text fragments; and mapping the multiple text fragments to the corresponding service types.

[0079] In this embodiment, during the process of determining the service type of a text fragment, if the service text is obtained, it can be segmented based on the dialogue rounds of the first and second objects to obtain multiple text fragments. These multiple text fragments are then mapped to their corresponding service types.

[0080] Optionally, a complete service text is received, generated from a speech-to-text dialogue, containing all textual records of multi-turn interactions between the first party (e.g., a sales consultant) and the second party (e.g., a customer). Basic cleaning of this service text can be performed, including removing redundant symbols, standardizing punctuation, merging broken sentences, and labeling speaker roles (e.g., "Sales:" "Customer:"), ensuring a clear text structure and semantic integrity, providing standardized input for subsequent round-by-round segmentation.

[0081] Optionally, based on the natural boundaries of speaker switching in the aforementioned service text, the complete service text is segmented according to dialogue rounds. Each round is defined as the smallest interactive unit in which a speaker speaks continuously until the other party responds. For example, "Salesperson: Hello, welcome" is the first round, the customer's reply "I'd like to see the XXX model on display" is the second round, the salesperson then says "Our X7 has a lot of space" is the third round, and so on. The system parses the speaker identifier sentence by sentence, merging consecutive statements from the same speaker into a single text segment, while statements from different speakers are segmented as independent rounds, forming a set of discrete text segments based on dialogue rounds. This segmentation method does not rely on a fixed number of words or punctuation, but is based on real interaction logic, ensuring that each segment carries a complete semantic action unit.

[0082] Optionally, a pre-trained lightweight classification agent is invoked, with each text fragment as an independent input, fed into the classification agent (classification model). The training data for this classification agent comes from historically labeled sales dialogues, and the output label set is a predefined set of service types, such as "greeting upon arrival," "demand mining," "product explanation," "objection handling," "test drive invitation," "price negotiation," "closing the deal," and "casual conversation." The model performs probabilistic classification based on keywords, semantic patterns, and intent tendencies in the fragments, outputting the most likely service type label and confidence score for each text fragment.

[0083] Optionally, due to the ambiguity of language expression or errors in speech transcription, the classification model may misclassify or fluctuate in individual segments. For example, in the product explanation stage, the sentence "Are you here to see the car today?" might be incorrectly classified as "greeting upon arrival." In this case, a sequence smoothing algorithm is activated to perform temporal analysis on the service type labels of each text segment. For example, a sliding window voting mechanism can be used, with three rounds as the unit. If the classification result (service type) of the current segment is inconsistent with the mainstream types of the previous and following two rounds, and the confidence level is lower than the confidence threshold (e.g., 0.7), then the above classification result is corrected to the nearest majority type. If a certain service type appears only once and is isolated from other consecutive service types (e.g., "greeting upon arrival" appears in the middle of "product explanation"), then it is directly removed and merged into the context-dominant type. Through the above process, the sequence of service types is ensured to conform to the actual business logic, eliminating jumps and noise.

[0084] Optionally, after the above processing, each text fragment is assigned a stable, coherent service type label that conforms to the business process.

[0085] In this embodiment, the above method achieves a shift from blind inspection of the entire text to precise segmentation, aligning quality inspection behavior with the business stages of the service process and significantly improving the accuracy and semantic relevance of rule matching. Segmentation based on dialogue rounds rather than fixed lengths better reflects real-world interaction semantics and avoids semantic fragmentation caused by mechanical truncation. Through a dual mechanism of intelligent classification and sequence smoothing, the method effectively suppresses classification jitter caused by contextual ambiguity in long texts by large models, significantly improving the stability and continuity of service type classification.

[0086] As an optional embodiment, mapping multiple text fragments to corresponding service types includes: invoking a classification agent to determine the service stages to which the multiple text fragments belong; in response to multiple service stages corresponding to consecutive multiple text fragments being in an abnormal state, using the classification agent to adjust at least one of the consecutive multiple service stages to obtain an adjusted service stage, wherein the adjusted consecutive multiple service stages are in a normal state; and mapping the adjusted service stage to a service type.

[0087] In this embodiment, during the process of mapping multiple text fragments to corresponding service types, a classification agent can be invoked to determine the service stage to which each text fragment belongs. If multiple service stages corresponding to consecutive text fragments are in an abnormal state, the classification agent can be used to adjust at least one of the consecutive service stages to obtain an adjusted service stage. The adjusted service stage is then mapped to a service type. Here, a service stage can refer to a continuous link in the service interaction process, defined by business logic and the customer journey, possessing clear objectives, typical behavioral patterns, and semantic features.

[0088] In this embodiment, service stages are equivalent to business stages, such as "greeting upon arrival," "demand mining," "product explanation," "objection handling," and "test drive invitation." Each service stage represents a standardized behavioral node in the interaction between the customer and the sales consultant, exhibiting sequential dependency and semantic coherence. The aforementioned abnormal state refers to the time-series jumps or discontinuous fluctuations that do not conform to the actual business logic after multiple consecutive text segments are mapped to service stages. A typical manifestation is the appearance of logically impossible or highly unreasonable stage jumps within a short period, such as "product explanation" → "greeting upon arrival" → "product explanation" → "product explanation," where "greeting upon arrival," as the initial stage of the service, reappears during the product explanation process, clearly violating the unidirectional evolution of the service process. This anomaly is usually caused by speech-to-text errors, semantic ambiguity, insufficient model classification confidence, or missing contextual information, and is not a reflection of actual service behavior. The essence of the abnormal state is "noise signal," which disrupts the integrity of the service process. If not corrected, it can lead to subsequent mis-triggered quality inspection rules, invalid arbitration logic, and distorted management insights.

[0089] Optionally, the system receives multiple text fragments segmented from dialogue rounds, each representing a complete unit of expression by a speaker (salesperson or customer). A pre-built classification agent is invoked; this agent is a domain-tuned language model trained on a large number of labeled car sales dialogues, with a predefined set of service stages as the label set. The classification agent analyzes the semantic content of each text fragment, identifying key intent words, action verbs, and contextual cues, and outputs a service stage label and its confidence score for each fragment.

[0090] Optionally, the output service stage labels are sequentially scanned in chronological order to identify any non-logical jumps. The criteria for judging abnormal states are based on the unidirectional evolution of the business process: normal service stages should exhibit a stable progression of "preceding → subsequent," for example, "greeting upon arrival" should be followed by "demand mining," and then "product explanation." There should be no atypical paths such as suddenly jumping back to "greeting upon arrival" after "product explanation" or repeatedly engaging in "casual conversation" after "objection handling." The system uses a sliding window mechanism, checking for patterns such as "ABA" or "ABCBA" in units of three to five consecutive segments. If a service stage appears as an isolated node within a short period, and its preceding and following segments belong to other stages, it is judged as an abnormal state. Simultaneously, if the classification confidence of a certain stage is lower than a preset threshold (e.g., 0.6), and its contextual semantics support other stages, it is also included in the abnormal candidate list.

[0091] Optionally, when an abnormal state is detected, the default correction rule is not immediately applied. Instead, the classification agent is reactivated. This time, the input is not a single text fragment, but a context window consisting of at least three fragments before and after the abnormal text fragment. For example, if text fragment 5 is determined to be "greeting upon arrival," but text fragments 4 and 6 are both "product explanation," then seven fragments (2 through 8) are submitted as new input to the classification agent. In this case, the agent no longer relies solely on the semantics of a single sentence, but infers based on the overall dialogue context: the sudden appearance of a "welcome" phrase during the "product explanation" stage is highly likely a residual phrase mis-segmented when the customer first entered the showroom, or a speech-to-text error. Based on contextual consistency, the classification agent corrects the isolated fragment to "product explanation" and updates its label and confidence level.

[0092] Optionally, after adjustment, a second verification is performed on all corrected service stage sequences to ensure that abnormal states have been eliminated and the entire sequence meets the temporal rationality of the business process. If multiple anomalies still exist, a third stage can be iteratively executed until the sequence stabilizes, or a conservative mode can be entered, retaining segments that are difficult to judge as "undetermined" and handing them over to the subsequent arbitration module for indirect correction based on quality inspection results. This stage ensures that the final mapping result conforms to the standard process model of sales and service at the macro level. The adjusted stable service stage label is used as the final service type mapping result. Service stage and service type are synonymous concepts in this system, so the adjusted label is directly used as the output to form a structured set of text segments, each segment with the finally confirmed service type, original text, confidence level, and whether it has been corrected. This set will be passed to the next stage—the concurrent quality inspection module—as the context for rule matching.

[0093] In this embodiment, the above method makes service type mapping no longer dependent on the randomness of single-sentence judgment, but intelligently corrects itself based on the overall logic of the dialogue, significantly reducing misjudgments caused by transcription errors or semantic ambiguity; through the dynamic context window re-judgment mechanism, the rigidity of traditional hard-coded rules (such as "no back") is avoided, effectively eliminating "timing jitter" and "logical jump" in the service stage sequence, ensuring that the context relied upon by the subsequent quality inspection agent is pure, continuous, and semantically consistent, greatly improving the rule hit accuracy and arbitration effectiveness; fourth, this process does not require manual rule intervention or additional training, and can achieve self-correction only through iterative calls of model inference, with engineering advantages of low operation and maintenance costs and high generalization ability.

[0094] As an optional embodiment, the method further includes: acquiring dialogue text characters in a text fragment; and determining a detection strategy based on the number and service type of the dialogue text characters.

[0095] In this embodiment, dialogue text characters in a text fragment can be acquired. A detection strategy is determined based on the number and service type of the dialogue text characters. These dialogue text characters can be language units, i.e., tokens, that have undergone word segmentation and encoding processing; they can be basic semantic units used by a natural language processing system during text analysis, model input, and computation.

[0096] Optionally, for each text fragment of a mapped service type, a standardized tokenizer and encoder (such as SentencePiece or BPE) is first invoked to convert the original text of the fragment into a sequence of tokens. The system counts the total number of tokens in this sequence, which is taken as the number of dialogue text characters for that text fragment. An internal "strategy mapping table" is pre-configured, binding each service type to one or more token threshold ranges and corresponding to different detection strategies. If the text length exceeds 3000 tokens, it automatically upgrades to a "fragmented retrieval + local weighting" strategy. This mapping table is developed based on business experience and historical data analysis to ensure that the strategy selection matches the semantic complexity of the service stage.

[0097] Optionally, the number of tokens in the current text fragment and its service type are jointly determined, matched against a policy mapping table, and the corresponding detection policy is executed. For "full context detection," the entire text fragment is used as complete input and directly fed into the quality control agent, forcing it to output thought chains and quoted original text, suitable for short texts and high semantic density scenarios. For "conditional RAG detection," the service type of the fragment and the current rule to be inspected (e.g., "whether a competitor's name is mentioned") are used as a joint query. Similar historical fragments are retrieved from a vector database, and the Top-K most relevant dialogue content is extracted. Only these relevant fragments are input into the quality control agent along with the original rule to avoid interference from irrelevant information. For the "fragmented retrieval + local weighting" strategy, the extremely long text is sliced ​​into fixed lengths (e.g., 500 tokens), each slice is independently retrieved and its relevance weight is calculated, and then the results are fused according to the weights to ensure that long texts do not lose key clues.

[0098] Optionally, regardless of the detection strategy employed, the quality inspection agent can be required to include a "MatchedQuote" field in its output, which is a quote from the original text that was determined to be matched. This quote is aligned with the input search content or original fragment to ensure the detection behavior remains within the context and avoids misjudgments caused by irrelevant statements introduced by RAG retrieval. If the quoted content in the detection result is inconsistent with the input fragment, the system records the anomaly and triggers a log alert, serving as a basis for subsequent model optimization. Information such as the strategy type, number of tokens, service type, number of search fragments, whether RAG was enabled, and the final quality inspection conclusion are all recorded as metadata and bound to the quality inspection results to form a complete detection log. This log is not only used for result traceability but also provides data support for subsequent strategy tuning, cost analysis, and model feedback loop closure.

[0099] In this embodiment, the above method enables adaptive dynamic switching of the detection strategy, abandoning the traditional "one-size-fits-all" full input mode. It intelligently selects the optimal path based on text length and business stage, significantly reducing token consumption and inference latency. Through the joint decision-making mechanism of token quantity and service type, it has the ability to process dialogues of different complexities, avoiding over-computation in short texts and hasty truncation in long texts, thus comprehensively improving detection accuracy.

[0100] As an optional embodiment, the detection strategy includes a first detection strategy and a second detection strategy. The first detection strategy represents the rules for detecting individual characters in the dialogue text, and the second detection strategy represents the rules for detecting the entire text segment. Step S204 involves calling the detection strategy corresponding to the service type to detect the text segment and obtain a first detection sub-result for the text segment. This includes: responding to a situation where the number of dialogue text characters is greater than or equal to a quantity threshold, calling the first detection strategy of the quality inspection agent to perform switching and vectorization processing on the dialogue text characters to obtain a processing result; calling the first detection result and determining the first detection sub-result based on the processing result; and responding to a situation where the number of dialogue text characters is less than the quantity threshold, calling the second detection strategy of the quality inspection agent to directly detect the text segment to obtain the first detection sub-result.

[0101] In this embodiment, during the detection of text fragments using a detection strategy, if the number of dialogue text characters is greater than or equal to a threshold, the first detection strategy of the quality inspection agent can be invoked to perform switching and vectorization processing on the dialogue text characters, obtaining a processing result. The first detection result can be invoked, and a first detection sub-result can be determined based on the processing result. If the number of dialogue text characters is less than the threshold, the second detection strategy of the quality inspection agent can be invoked to directly detect the text fragment itself, obtaining a first detection sub-result. The first detection strategy can refer to a detection method for long text fragments based on retrieval enhancement and local focusing. When the number of dialogue text characters (Tokens) reaches or exceeds a preset threshold (e.g., 3000), the first detection strategy is adopted. The core idea is not to input the entire text fragment into a large model at once, but to segment the text fragment, converting each segment into a vector representation through an embedding model, using the current rule definition to be detected as the query statement. The second detection strategy can refer to a direct detection method for short text fragments based on full context understanding. The core of the second detection strategy mentioned above is to input the entire text segment into the quality inspection agent, which then makes a one-time judgment on the entire text segment based on its built-in semantic understanding capabilities and rule definitions.

[0102] Optionally, the quantity threshold can refer to a critical value used to determine whether the text length enters "long text mode". In this embodiment, it can be set to 3000 tokens. The above processing result can refer to the product generated after text segmentation, vectorization and vector retrieval when using the first detection strategy.

[0103] Optionally, upon receiving a text fragment mapped by service type, a standardized text encoder is invoked to convert the original dialogue text into a sequence of tokens, and the total number of tokens is precisely counted. This number serves as the sole objective indicator of the text information volume and is the sole basis for subsequent strategy selection. When the number of tokens is greater than or equal to 3000, the system initiates the first detection strategy. This strategy does not directly input the full text but performs two key processes: first, switching processing, which logically segments the long text into multiple semantically coherent sub-fragments by a fixed length (e.g., 500–800 tokens); second, vectorization processing, which uses a lightweight embedding model to encode each sub-fragment into a high-dimensional dense vector and stores it in a vector database. The current quality inspection rule to be detected (e.g., "whether the competitor's name is actively mentioned") is used as a query statement, also vectorized, and a similarity search is performed in the database to extract the Top-K text fragments that are semantically closest to the rule, forming a highly focused "relevant context subset".

[0104] Optionally, the Top-K relevant fragments obtained in the previous step are input into the quality inspection agent along with the original quality inspection rules, and the agent is forced to output a "thought chain" and "quoted original text". The quality inspection agent then reasons based solely on selected fragments, rather than dealing with thousands of words of redundant content, thus avoiding information overload and attention drift. When the number of tokens is less than 3000, the second detection strategy is activated. This second detection strategy directly inputs the entire text fragment as complete data into the quality inspection agent without any segmentation or retrieval. At this point, the model can fully read the dialogue context, identifying subtle features such as semantic connections, tone changes, and implicit intentions. This is particularly suitable for rules requiring an understanding of the overall context, such as "whether the demand mining loop is complete" or "whether a test drive is guided at the appropriate time". The system also forces the agent to output a thought chain and quoted original text, ensuring the reasoning process is transparent and traceable. The complete quality inspection conclusion output is the first detection sub-result of the text fragment.

[0105] Optionally, regardless of the strategy used to generate the results, they will all be uniformly formatted as structured JSON output, including metadata fields such as detection conclusion, original text, confidence level, strategy type (first or second), processing time, token consumption, and service type, and recorded in the system log.

[0106] In the embodiments of this application, the above method realizes intelligent divide-and-conquer by using retrieval for long texts and full data collection for short texts, solving the dilemma of related technologies being either too costly or lacking in accuracy; the first detection strategy accurately focuses on relevant semantics through rule-driven retrieval, while the second detection strategy retains the model's advantage of global understanding of short texts, avoiding over-processing of simple dialogues and ensuring maximum detection efficiency.

[0107] As an optional embodiment, the method further includes: invoking a review agent to review the first detection sub-result and obtain a review result, wherein the review result is used to represent the detection quality of the first detection sub-result; and in response to the detection quality being lower than a quality threshold, determining that the accuracy of the first detection sub-result is less than an accuracy threshold.

[0108] In this embodiment, a review agent can be invoked to review the first detection sub-result and obtain a review structure. If the detection quality is lower than a quality threshold, it can be determined that the accuracy of the first detection sub-result is less than the accuracy threshold. The review result can be a structured judgment conclusion generated by the review agent after logically verifying the consistency and compliance of the first detection sub-result output by the quality inspection agent. The function of the review result is to identify and mark potential hidden logical errors or semantic deviations that may exist during the initial inspection process.

[0109] Optionally, after obtaining the first detection sub-result output by the quality inspection intelligent agent, the system automatically determines whether the result meets the preconditions for entering the review process. These preconditions do not rely on manual intervention but are automatically activated based on preset rule complexity and risk levels. For example, when the first detection sub-result involves rules with subjective intent judgments such as "actively provided," "whether promised," or "whether induced," a potential semantic trap is identified, and the review process is automatically initiated. If it is a simple keyword matching rule (such as "whether 'Welcome' appears"), the review is skipped, and the process directly enters the aggregation stage. This step ensures that review resources are only used for high-risk, high-value detection results, avoiding computational redundancy.

[0110] Optionally, the entire content of the first detection sub-result, along with the original quality inspection rule definition, the original dialogue fragment, and the context service type information, is uniformly assembled into the input context of the review agent. This input package is highly constrained: the rule definition clarifies "what should be detected," the original text fragment provides the basis for judgment, the service type is used to calibrate contextual expectations, and the first detection sub-result serves as a candidate conclusion to be challenged. The review agent performs deep semantic analysis based on the input package and outputs judgments using a structured thought chain. Its core actions include: comparing whether the rule keywords are semantically consistent with the actual quoted text; for example, the rule requires "the salesperson proactively asks if I need a drink," but the initial detection result only quotes "the salesperson pours water" without mentioning "whether they asked"; judging whether the initiator of the behavior is incorrectly attributed; for example, misjudging "Can the customer pour me a glass of water?" as "the salesperson proactively offered it"; and identifying distortions of implicit intent; for example, understanding "the price is too high" as "the customer has the intention to buy" instead of "the objection has not been resolved."

[0111] Optionally, the "has_problem" field in the review result serves as a direct indicator of detection quality, with a Boolean value: if false, it indicates that no logical flaws were found during the review, and the detection quality meets the standard; if true, it indicates that the detection result has semantic deviations or rule misinterpretations, and the detection quality does not meet the standard. The quality threshold here is an implicit setting, namely, "whether there are any explicitly identifiable logical errors." The "has_problem" status in the review result is used as a decision signal to directly link subsequent processes: if false, the first detection sub-result is marked as "passed review" and enters the result aggregation stage; if true, the system automatically marks the detection result as questionable and transmits it to the arbitration agent, along with review comments as a basis for correction, awaiting final adjudication.

[0112] As an optional embodiment, step S206, in response to the first detection sub-result indicating that the accuracy of the performance index of the detection specification is less than the accuracy threshold, detects the text segment to obtain a second detection sub-result of the text segment, including: in response to the first detection sub-result indicating that the accuracy of the performance index of the detection specification is less than the accuracy threshold, calling the review agent to review the text segment and obtain the second detection sub-result.

[0113] In this embodiment, during the process of detecting text fragments and obtaining the second detection sub-result, if the accuracy of the detection specification performance index indicated by the first detection sub-result is less than the accuracy threshold, the review agent can be invoked to review the text fragments and obtain the second detection sub-result.

[0114] Optionally, after generating and reviewing the first detection sub-result, the system determines whether the detection quality is below a quality threshold based on the review result output by the review agent. If "has_problem" is true in the review result, it indicates that the first detection sub-result has logical flaws, semantic misinterpretations, or rule violations. Based on this, the system determines that the accuracy of its specification performance indicators is below a preset accuracy threshold. This threshold represents the minimum acceptable judgment baseline in the business scenario, such as a 90% semantic accuracy rate. Once it falls below this threshold, the system determines that the result cannot be used as the final basis and must enter the correction process. Instead of relying on the conclusion of the first detection sub-result, the original input—the complete dialogue text fragment, the original quality inspection rule definition, and the service type label—is reloaded into the input context of the review agent.

[0115] Optionally, the review agent re-analyzes the text fragments from a new perspective, performing complete semantic understanding, intent judgment, and fact-checking. Its reasoning process still mandates the output of thought chains, ensuring that each judgment is supported by the original text. The generated second detection sub-result undergoes two checks: first, logical consistency, confirming that its cited original text completely matches the rule definition, without subjective expansion or omission; second, the uniqueness of the conclusion, ensuring that it is not cited or influenced by the first detection sub-result, achieving truly independent judgment. If the checks pass, the result is locked as valid output; if ambiguity or contradiction is still found, the system records an anomaly log, entering the manual review channel as feedback data for model iteration.

[0116] Optionally, the second detection sub-result automatically replaces the first detection sub-result, serving as the sole input to the subsequent aggregation and insight modules. While the first detection sub-result is retained in the logs for auditing purposes, it is no longer involved in the final report generation. The entire correction process is completed in milliseconds, requiring no manual intervention or re-calling of other models; the result correction is achieved solely through the review agent's permission upgrade and inference reset.

[0117] As an optional embodiment, the method further includes: invoking an insight agent to generate a management policy for a first object based on the detection results, wherein the management policy is used to represent rules for managing the first object to improve the standard performance indicators of the first object in providing services to the second object.

[0118] In this embodiment, an insight agent can be invoked to generate a management strategy for the first object based on the detection results. This management strategy can be a macro-level recommendation and summary insight generated by the insight agent based on a large amount of quality inspection results, geared towards management decision-making. Essentially, this management strategy is high-level business guidance information output after semantic aggregation, pattern recognition, and value extraction of dispersed, micro-level detection data.

[0119] Optionally, the intelligent agent performs statistical and semantic clustering analysis on the aggregated detection results to identify three core anomaly patterns: The first is "high-frequency misses," which are rules repeatedly not met in multiple dialogues, such as "not proactively asking about the budget" appearing 42 times in 50 recordings; the second is "common deviations across multiple journeys," where the same type of problem appears in different service stages, such as "avoiding price issues" appearing in both the product explanation stage and the objection handling stage; the third is "anomaly clusters," where the violation rate in a certain service journey is significantly higher than in other stages, such as the miss rate in the "test drive invitation" stage being three times that of other stages. This stage does not rely on manually defined rules, but rather the intelligent agent autonomously discovers statistically significant deviations to identify the true systemic shortcomings.

[0120] Optionally, after identifying anomalous patterns, the insight agent further deduces the underlying behavioral motivations and management causes. It goes beyond simply stating "the salesperson didn't ask about the budget," analyzing the context of the conversation to determine "why they didn't ask." For example, if multiple missed records show that the salesperson skipped asking about the budget after the customer remained silent, the agent infers that "lack of guiding communication" or "fear of customer backlash" are potential psychological barriers. If the salesperson uses vague responses like "our prices are quite reasonable" in multiple conversations, the agent judges that "lack of standardized pricing communication techniques" is a systemic deficiency. By mapping specific statements to behavioral intentions or training deficiencies, the insight agent elevates "detected facts" to "management problems."

[0121] Optionally, based on attribution analysis, the insight agent invokes a preset policy generation template and outputs a management policy in natural language.

[0122] The technical solutions of the embodiments of this application will be illustrated below with reference to preferred embodiments.

[0123] Currently, in the business model between automakers and dealers, rebates based on vehicle sales are a common incentive. In existing technology, business departments (such as marketing and sales departments) generally rely on spreadsheet software (such as Excel) for manual calculation of rebate amounts. This process typically involves exporting data from multiple heterogeneous business systems (such as sales systems, financial systems, and inventory systems), and then manually screening, matching, filtering, aggregating, and applying complex rebate policy formulas for calculation.

[0124] Currently, related technical solutions with similar objectives to the embodiments of this application mainly include the following categories:

[0125] Quality checks, whether manual or rule-based, rely on manual sampling or regular expression matching of predefined keywords (such as "Welcome" or "Add me on WeChat"). These methods fail to understand semantic context (e.g., whether "Add me on WeChat" is initiated by the salesperson or the customer), resulting in a very high false positive rate.

[0126] General quality inspection services based on cloud vendor APIs directly call the quality inspection APIs of third-party cloud services. These services are usually charged by time and are expensive (approximately 0.13 yuan / hour or more). At the same time, their quality inspection logic is usually a black box, making it difficult to deeply customize and arbitrate logic for specific business journeys (such as "demand mining" and "test drive invitation" specific to car sales).

[0127] For single-unit large-scale model quality checks, the entire long text is directly input into the large-scale model (LLM) for a one-time evaluation. Since car sales recordings are usually long (thousands of words), a single input can easily lead to model attention being distracted (Lost in the Middle) and is prone to illusions, making it difficult to accurately locate subtle logical errors.

[0128] The aforementioned technologies suffer from the following problems: Logical judgments may be "factually correct but with a flawed intent," and large-scale model quality control solutions (monomeric language models) often only perform shallow semantic matching. When faced with scenarios where the actions ("customer actively requests a free gift") and "salesperson actively provides a free gift") are consistent in fact but have diametrically opposed subjective intentions, monomeric models are prone to misjudgment due to misinterpretations. There is also the paradox of "lost in the middle" and high costs, as transcripts from long service scenarios like car sales often exceed several thousand words. Direct full input leads to model "lost in the middle" and exponentially increases token costs; simple truncation results in the loss of crucial cross-paragraph context; there is a lack of a closed-loop automated error correction mechanism: existing quality control systems typically rely on manual appeals or subsequent model iterations to correct errors, lacking an automated arbitration and self-correction mechanism that takes effect immediately during the online inference phase; and there is a lack of management-perspective insights: existing outputs are mostly fragmented score sheets, lacking aggregated analysis and suggestion generation for overall sales strategies and macro-level issues.

[0129] The embodiments of this application will be further described below.

[0130] This application aims to address the technical problems of high cost, large misunderstanding of long texts, high misjudgment rate of complex logic, and lack of in-depth management insight in existing dialogue quality inspection technologies. It provides an offline dialogue quality inspection method and system based on multi-agent hierarchical collaboration, with self-arbitration and correction capabilities, and extremely low cost.

[0131] Compared to mainstream cloud vendors' quality inspection services (approximately RMB 0.13 / hour), the technical solution described in this application reduces the cost to approximately RMB 0.0084 / thousand tokens (approximately 1 / 15th of the competitor's overall cost) through self-developed Prompt engineering and Token optimization, significantly reducing the operational burden on enterprises. A highly accurate arbitration and error correction mechanism introduces a multi-level intelligent agent architecture of "quality inspector-reviewer-arbitrator." Through "AI checking AI," it can automatically identify and correct complex logical errors such as "customer actively requesting information being misjudged as sales actively providing it," significantly reducing the illusion rate. Precise processing of long texts and context, combined with conditional RAG (retrieval augmentation generation) technology, automatically enables segmented retrieval for dialogues exceeding 3000 words, ensuring accurate fact-checking without losing context. Accurate reconstruction of business processes, through the dual guarantee of "intelligent agent classification + code smoothing algorithm," accurately segments business journeys such as "demand mining" and "product explanation," solving the noise problem of classification results jumping across the timeline (such as ABAA). It can automatically generate in-depth insight reports from a manager's perspective, transforming dry data into actionable improvement suggestions (such as SOP script optimization directions).

[0132] Figure 3 This is a schematic diagram of an offline dialogue quality inspection system based on a multi-agent collaboration and arbitration mechanism, according to an embodiment of this application. Figure 3 As shown, the core concept of the offline dialogue quality inspection system is "journey segmentation, concurrent quality inspection, multi-layer review, and insight generation". The offline dialogue quality inspection system 300 mainly comprises the following six core processing modules:

[0133] The task receiving and rule distribution module 302 is used to receive quality inspection requests, parse the quality inspection rules configured in the business (the rules include titles, definitions, positive / negative standards), and intelligently distribute the rules to the "global quality inspection" or "specific journey quality inspection" queue.

[0134] The intelligent dialogue journey segmentation module 304 uses a "Journey Classifier Agent" combined with a "sequence smoothing algorithm" to map linear dialogue text to specific business journeys (such as in-store greeting, demand mining, etc.).

[0135] The concurrent quality inspection and conditional RAG module 306 utilizes a "Quality Inspection Agent" to execute quality inspection tasks in parallel. For extremely long texts (e.g., >3000 characters), it dynamically triggers the RAG (Retrieval Enhancement Generation) mechanism, using embedding to retrieve relevant segments to assist in quality inspection.

[0136] The multi-round review and arbitration module 308 introduces a "Reviewer" to conduct a logical review of the initial inspection results. Once a doubt is found, the "Decider" is triggered to make a final judgment and correction.

[0137] The results aggregation module 310 provides a structured summary of all journey quality inspection results after arbitration correction.

[0138] AI Manager Insight Generation Module 312 generates macro-level management suggestions and summaries based on aggregated results using the "Insight Agent".

[0139] Optionally, during task reception and rule distribution, the backend reads a pre-configured rule set. The code logic categorizes rules based on their attributes (Global vs. Journey-specific). The filtered rule set is then dynamically injected into the System Prompt of subsequent agents, enabling targeted quality checks and preventing aimless wandering of general large models.

[0140] Optionally, during intelligent journey segmentation, the complete dialogue text is physically segmented into one-turn segments. Lightweight models such as qwen-turbo are used as "classification agents" to concurrently determine the business stage to which each segment belongs (e.g., customer greeting, needs assessment, objection handling, small talk). A built-in "sequence smoothing algorithm" is used for post-processing to address potential jitter in model classification (e.g., A->B->A->A). If a non-logical jump is detected within a short period (e.g., a misclassified "customer greeting" is suddenly inserted into a deep "product explanation"), the algorithm corrects the jump to a context-consistent category, ensuring the continuity of the journey.

[0141] Optionally, during concurrent quality checks and conditional RAG, a "global agent" is activated to monitor the entire process (e.g., whether profanity is used), while multiple "journey agents" are activated to focus on specific stages (e.g., in the "demand mining" stage, only checking whether the budget was asked). The Prompt forces the model to output its thought process and original text citations, prohibiting direct output of results to reduce illusions. If the dialogue text tokens < 3000, the entire context is directly input. If tokens > 3000, the text is sliced ​​and embedded. Quality check rules are used as queries for vector retrieval to extract Top-K relevant fragments. Only relevant fragments are input into the quality check agent, which significantly improves accuracy and reduces token consumption for "fact-checking" rules (e.g., whether specific competitor names are mentioned).

[0142] Optionally, the multi-round review and arbitration process simulates the "verification" process of human quality inspection, which is key to ensuring high accuracy. Role A is the Quality Inspector (Executor), producing preliminary quality inspection results (including hit status and citations). Role B is the Reviewer (Checker), used to "find faults." Their Prompt is designed to find logical loopholes. For example, when the rule requires "sales to proactively offer drinks," and the initial check result shows a hit, but the original quote is "Customer: Can you pour me a glass of water?", the Reviewer will mark this entry as has_problem: true, pointing out that "customer request does not equal sales proactive." Role C is the Arbitrator (Decider): intervenes only when the Reviewer finds a problem. They receive the original text, the initial check conclusion, and the reviewer's comments, make the final weighing, and output the final JSON result. This mechanism can automatically correct a large number of hidden logical errors.

[0143] Optionally, during the result aggregation and insight generation process, the backend service collects all "approved" initial inspection results and "corrected" arbitration results, merging them into a final quality inspection report. The "AI Manager Agent" receives the final report, ignoring specific scoring details, and instead focuses on high-frequency "missed" items and negative labels, generating a Markdown-formatted "Execution Summary and Improvement Suggestions" that points out the sales team's macro-level shortcomings in dimensions such as "proactiveness" and "objection handling skills."

[0144] According to an embodiment of this application, a vehicle service text processing apparatus is also provided. It should be noted that this vehicle service text processing apparatus can be used to execute the vehicle service text processing method described in the above embodiments.

[0145] Figure 4 This is a schematic diagram of a vehicle service text processing apparatus according to an embodiment of this application, such as... Figure 4 As shown, the service text processing device 400 for the vehicle may include: a first determining unit 402, a first detecting unit 404, a second detecting unit 406, and a second determining unit 408.

[0146] The first determining unit 402 is used to determine the service type of the text fragment in response to obtaining the service text, wherein different service types correspond to different service contents of the service provided by the first object to the second object.

[0147] The first detection unit 404 is used to call the detection strategy corresponding to the service type to detect the text fragment and obtain the first detection sub-result of the text fragment. The detection strategy is used to represent the rules for detecting the standard performance indicators of the service content in the text fragment.

[0148] The second detection unit 406 is used to detect the text segment in response to the first detection sub-result indicating that the accuracy of the performance index is less than the accuracy threshold, and to obtain a second detection sub-result of the text segment, wherein the second detection sub-result indicates that the accuracy of the performance index is greater than or equal to the accuracy threshold.

[0149] The second determining unit 408 is used to determine the detection result of the service text based on multiple second detection sub-results corresponding to multiple text fragments, wherein the detection result is used to represent the standard performance index of the service provided by the first object to the second object.

[0150] In this embodiment, the first determining unit 402, in response to obtaining the service text, determines the service type of the text fragment, wherein different service types correspond to different service contents provided by the first object to the second object; the first detection unit 404 calls the detection strategy corresponding to the service type to detect the text fragment, obtaining a first detection sub-result of the text fragment, wherein the detection strategy is used to represent the rules for detecting the standard performance indicators of the service content in the text fragment; the second detection unit 406, in response to the first detection sub-result indicating that the accuracy of the standard performance indicator is less than the accuracy threshold, detects the text fragment, obtaining a second detection sub-result of the text fragment, wherein the second detection sub-result indicates that the accuracy of the standard performance indicator is greater than or equal to the accuracy threshold; the second determining unit 408, based on multiple second detection sub-results corresponding to multiple text fragments, determines the detection result of the service text, wherein the detection result is used to represent the standard performance indicators of the service provided by the first object to the second object, thereby solving the technical problem of low processing accuracy of vehicle service text and achieving the technical effect of improving the processing accuracy of vehicle service text.

[0151] According to an embodiment of this application, a computer-readable storage medium is also provided, the storage medium including a stored program, wherein the program executes the methods described in the embodiments of this application.

[0152] According to an embodiment of this application, a processor is also provided for running a program, wherein the program executes the methods described in the embodiments of this application during runtime.

[0153] According to another aspect of the embodiments of this application, Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application, such as... Figure 5 Furthermore, an electronic device 50 is also provided. This electronic device 50 includes a memory 501 and a processor 502. The memory 501 stores a computer program, and the processor 502 is configured to run the computer program to perform the methods described in the embodiments of this application.

[0154] Embodiments of this application may provide a computing device. Figure 6 This is a structural block diagram of a computing device according to an embodiment of this application. Figure 6 As shown, the computing device 600 may include: one or more (one shown in the figure) processors 602, memory 604, memory controller, and peripheral interfaces.

[0155] The aforementioned computing device can be understood as an integrated intelligent terminal, including but not limited to servers, desktop computers, personal computers (PCs), and all-in-one model machines. Furthermore, the computing device may be pre-installed with a model for executing the vehicle service text processing method described in the above embodiments of this application.

[0156] According to another aspect of the embodiments of this application, a computer program product is also provided. This computer program product includes a computer program that, when executed by a processor, implements the methods described in the embodiments of this application.

[0157] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0161] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0162] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing service text for vehicles, characterized in that, The service text is used to represent the process by which a first object provides a service for the vehicle to a second object. The service text includes multiple text fragments. The method includes: In response to obtaining the service text, the service type of the text fragment is determined, wherein different service types correspond to different service contents of the service provided by the first object to the second object; The detection strategy corresponding to the service type is invoked to detect the text fragment, and a first detection sub-result of the text fragment is obtained. The detection strategy is used to represent the rules for detecting the standard performance indicators of the service content in the text fragment. In response to the first detection sub-result indicating that the accuracy of the standard performance index is less than the accuracy threshold, the text segment is detected to obtain a second detection sub-result of the text segment, wherein the second detection sub-result indicates that the accuracy of the standard performance index is greater than or equal to the accuracy threshold; Based on multiple second detection sub-results corresponding to multiple text fragments, a detection result of the service text is determined, wherein the detection result is used to represent the canonical performance index of the service provided by the first object to the second object.

2. The method according to claim 1, characterized in that, In response to obtaining the service text, determining the service type of the text fragment includes: In response to obtaining the service text, the service text is segmented based on the dialogue rounds between the first object and the second object to obtain multiple text fragments; Map multiple text fragments to their corresponding service types.

3. The method according to claim 2, characterized in that, Mapping multiple text fragments to corresponding service types includes: The classification agent is invoked to determine the service stage to which the multiple text fragments belong; In response to multiple service stages corresponding to a series of consecutive text segments being in an abnormal state, the classification agent is used to adjust at least one of the consecutive service stages to obtain the adjusted service stage, wherein the adjusted consecutive service stages are in a normal state. The adjusted service stage is mapped to the service type.

4. The method according to claim 1, characterized in that, The method further includes: Extract the dialogue text characters from the text fragment; The detection strategy is determined based on the number of characters in the dialogue text and the service type.

5. The method according to claim 4, characterized in that, The detection strategy includes a first detection strategy and a second detection strategy. The first detection strategy represents the rules for detecting individual characters in the dialogue text, and the second detection strategy represents the rules for detecting the entire text segment. The detection strategy corresponding to the service type is invoked to detect the text segment, resulting in a first detection sub-result for the text segment, including: In response to the number of dialogue text characters being greater than or equal to a quantity threshold, the first detection strategy of the quality inspection agent is invoked to perform switching and vectorization processing on the dialogue text characters to obtain the processing result. The first detection result is invoked, and the first detection sub-result is determined based on the processing result; In response to the fact that the number of characters in the dialogue text is less than the number threshold, the second detection strategy of the quality inspection agent is invoked to directly detect the text fragment and obtain the first detection sub-result.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The review agent is invoked to review the first detection sub-result and obtain a review result, wherein the review result is used to represent the detection quality of the first detection sub-result; In response to the detection quality being lower than a quality threshold, it is determined that the first detection sub-result indicates that the accuracy is less than the accuracy threshold.

7. The method according to claim 6, characterized in that, In response to the first detection sub-result indicating that the accuracy of the specification performance index is less than the accuracy threshold, the text segment is detected to obtain a second detection sub-result for the text segment, including: In response to the first detection sub-result indicating that the accuracy of detecting the standard performance index is less than the accuracy threshold, the review agent is invoked to review the text fragment and obtain the second detection sub-result.

8. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The insight agent is invoked to generate a management policy for the first object based on the detection results. The management policy represents the rules for managing the first object to improve the standard performance indicators of the first object in providing the service to the second object.

9. A device for processing service text for a vehicle, characterized in that, The service text is used to represent the process by which a first object provides a service for the vehicle to a second object, and the service text includes multiple text fragments. The device includes: The first determining unit is configured to determine the service type of the text fragment in response to obtaining the service text, wherein different service types correspond to different service contents of the service provided by the first object to the second object. The detection unit is used to call the detection strategy corresponding to the service type to detect the text fragment and obtain the first detection sub-result of the text fragment, wherein the detection strategy is used to represent the rules for detecting the standard performance indicators of the service content in the text fragment; An adjustment unit is configured to, in response to the first detection sub-result indicating that the accuracy of the standard performance index is less than an accuracy threshold, detect the text segment to obtain a second detection sub-result for the text segment, wherein the second detection sub-result indicates that the accuracy of the standard performance index is greater than or equal to the accuracy threshold. Based on multiple second detection sub-results corresponding to multiple text fragments, a detection result of the service text is determined, wherein the detection result is used to represent the canonical performance index of the service provided by the first object to the second object.

10. A processor, characterized in that, The processor is used to run a program, wherein the program, when running, performs the method according to any one of claims 1 to 8.

11. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 8.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 8.

Images