An offline large language model-based intelligent research and judgment and pre-litigation handling method and system for online car-hailing complaints

Abstract

A method and system for intelligent assessment and proactive handling of ride-hailing negative reviews based on an offline large language model, relating to the fields of data processing and artificial intelligence. The method includes: receiving a table file of ride-hailing review data and performing intelligent field mapping to unify heterogeneous data into standard fields; performing void detection on the review content to filter meaningless and invalid review records; loading a user-defined set of review rules and dynamically constructing model prompt words based on the text to be analyzed; feeding the prompt words into a locally deployed offline large language model to complete semantic reasoning and outputting a structured assessment result containing risk level, problem classification, and handling suggestions; automatically generating a standardized handling work order based on the assessment result and exporting a standard editable handling document. This invention achieves fully offline automated processing of the entire process from importing negative review data, AI assessment, work order generation to document export. Sensitive review data is stored locally throughout the process and not transmitted externally. It is suitable for government intranet office scenarios without external network access, allows for customizable negative review grading and classification rules, and significantly reduces the workload of manual handling.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent analysis and proactive handling of negative reviews of ride-hailing services based on an offline large language model. Specifically, it relates to a method and system for intelligent analysis and proactive handling of negative reviews of ride-hailing services based on an offline large language model. Background Technology

[0002] With the rapid development of the ride-hailing industry, passenger evaluation data on the quality of travel services is increasing. Transportation authorities need to conduct timely and effective analysis and handling of massive amounts of negative reviews in order to protect passenger safety and legitimate rights.

[0003] Currently, the processing of negative ride-hailing reviews mainly relies on manual reading and classification of each review. Personnel handling these reviews need to determine if violations exist, assess risk levels, and create processing orders based on the content. This method has the following technical drawbacks: First, manually reading through large amounts of negative review data is time-consuming and labor-intensive, resulting in low processing efficiency. Second, different personnel may have varying understandings of the review content, leading to inconsistent judgment standards and a high degree of subjectivity in the processing results. Third, negative review data involves personal information and travel trajectories of passengers and drivers, constituting sensitive travel data. Existing processing methods require uploading the data to cloud servers for AI analysis, posing a data leakage risk. Fourth, government intranet environments typically cannot access the external network, making existing cloud-based large-scale model-based solutions unusable in government intranet scenarios.

[0004] Most existing tools for handling negative reviews of ride-hailing services rely on large-scale online models. However, government intranet environments cannot access the external network, and manually identifying each negative review and creating processing work orders is a labor-intensive task. There is a lack of a processing solution that can be deployed offline, automatically analyzed by local artificial intelligence, and exported work orders with one click. Summary of the Invention

[0005] To address the aforementioned deficiencies in existing technologies, the present invention aims to provide a method and system for intelligent analysis and pre-emptive handling of ride-hailing negative reviews based on an offline large language model. This system enables offline processing, automatic analysis, and work order export of negative review data throughout the entire process, solving the technical problems of data leakage, low efficiency of manual handling, inability to customize rules, and fragmented processes in existing solutions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent judgment and pre-emptive handling of ride-hailing negative reviews based on an offline large language model, comprising the following steps: S1: Data import and intelligent field mapping step; receiving ride-hailing negative review evaluation data table file uploaded by users, performing multi-strategy intelligent matching on the file header fields, wherein the multi-strategy matching includes a three-level priority strategy of exact matching, whitespace-removing exact matching, and field inclusion fuzzy matching, uniformly mapping heterogeneous header fields to the system's preset standard field names, completing the standardization processing of heterogeneous data; S2: Filtering of hollow and invalid data in negative review content step Step 1: Perform pre-emptive hole detection on the imported negative review text to remove blank and meaningless reviews. The hole detection includes minimum effective character threshold determination and pure repetitive text recognition. The minimum effective character threshold is 4 characters. The pure repetitive text recognition uses continuous repetitive character matching rules to identify meaningless spam review text. Step 2: Custom review rule loading and dynamic prompt word construction steps: Load the user-defined ride-hailing regulatory rule configuration, and concatenate the regulatory rule content with the filtered negative review text to construct standardized model prompt words. The regulatory rules include negative review rating classification standards and corresponding handling suggestions. S4: Local Offline Large Language Model Artificial Intelligence Semantic Judgment Step; The constructed model prompt words are sent to the local offline large language model inference service for semantic analysis. The local offline large language model inference service runs offline on the local device. The inference temperature parameter is set to 0.3, the maximum output characters for a single inference is limited to 1024, and the inference timeout threshold is set to 120 seconds. The model outputs a structured judgment result containing risk level, problem type, and handling suggestions; S5: Structured Judgment Result Storage Step; The structured results output by the artificial intelligence judgment are written into the local embedded storage module for storage. The local embedded storage module stores all evaluation data and judgment results. The data is encrypted and stored locally throughout the process and is not transmitted to external networks. S6: Automatic generation of standardized negative review handling work orders; Based on the judgment results stored in the local embedded storage module, a pre-litigation handling work order containing evaluation information, judgment conclusions, and handling suggestions is automatically generated according to a preset work order template. S7: Batch export of work orders to editable handling work order files; The generated pre-litigation handling work orders are batch exported as editable document format files, allowing handling personnel to directly edit and modify them in the document for use in offline handling processes.

[0007] A ride-hailing negative review intelligent judgment and pre-emptive handling system based on an offline large language model includes: a data import mapping module (101), used to receive ride-hailing negative review evaluation data table files uploaded by users, perform multi-strategy intelligent matching on the file header fields, and uniformly map heterogeneous header fields to the system's preset standard field names; an evaluation data parsing unit (103), integrated into the data import mapping module (101), used to parse the data content in the table file and pass the parsing results to the subsequent processing module; an offline artificial intelligence judgment module (102), used to load custom regulatory rules and construct model prompt words, send the prompt words to the local offline large language model inference service (108) for semantic analysis, and receive the structured judgment results returned by the model; and a handling work order. The export unit (104) is used to automatically generate pre-emptive handling work orders according to the judgment results and preset work order templates, and export the work orders in batches as editable document format files; the internal communication service module (105) is used to provide data communication interfaces and page access services between various modules within the system; the rule configuration authorization module (106) is used to manage user-defined regulatory rule configurations and system usage authorization verification; the local embedded storage module (107) is used to store all evaluation data, judgment results and system configuration information, and the data is encrypted and stored locally throughout the process; the local offline large language model inference service (108) is used to run large language model inference offline on local devices, receive model prompts and output structured judgment results containing risk level, problem type and handling suggestions.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) It supports offline deployment on the intranet and can run all functional modules without connecting to the Internet; (2) Artificial intelligence automatically identifies the risk level and problem type of negative reviews, reducing the workload of manual screening; (3) It automatically generates standardized pre-litigation handling work orders based on the judgment results and exports them as editable document files with one click; (4) All evaluation data is encrypted and stored locally to ensure the information security of passengers and drivers. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the data import mapping process of the present invention; Figure 4 This is a schematic diagram of the artificial intelligence judgment and processing flow of the present invention.

[0010] Figure labeling explanations: 101 Data import mapping module; 102 Offline AI analysis module; 103 Evaluation data parsing unit; 104 Work order export unit; 105 Internal communication service module; 106 Rule configuration and authorization module; 107 Local embedded storage module; 108 Local offline large language model inference service; 201 User-uploaded evaluation data table file; 202 Intelligent matching of header fields; 203 Standard field mapping confirmation; 204 Hollow and invalid data filtering; 205 Regulatory rule loading; 206 Dynamic prompt word construction; 207 Local offline large language model inference... The following steps are implemented: 208 Structured analysis results storage; 209 Generation and export of pre-emptive handling work orders; 301 Exact matching; 302 Space-removing exact matching; 303 Field inclusion fuzzy matching; 304 Successful mapping processing; 305 Mapping failure prompts; 306 Standard field name library; 307 Mapping result confirmation interface; 401 Negative review text preprocessing; 402 Regulatory rule library reading; 403 Prompt word template concatenation; 404 Local offline large language model inference service; 405 Analysis result parsing; 406 Storage of analysis results into local embedded storage module; 407 Analysis exception handling and retry. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The offline ride-hailing negative review handling system corresponding to the present invention is a specific implementation product of the present invention.

[0012] Example 1: System Deployment and Data Import. This example provides a specific implementation process for a method of intelligent judgment and pre-emptive handling of ride-hailing negative reviews based on an offline large language model. The offline ride-hailing negative review pre-emptive handling system corresponding to this invention is built on a desktop application framework, integrates an interactive front-end interface, and provides an internal data communication interface through an internal communication service module (105). Users upload ride-hailing negative review evaluation data table files (201) through the data import mapping module (101). The evaluation data parsing unit (103) reads the table file content and extracts the list of table header fields. The system performs a three-level priority intelligent matching strategy on the table header fields: the first level performs exact matching (301), comparing the table header field names with the standard field name library (306) to ensure complete consistency; the second level performs exact matching (302), removing leading and trailing spaces and spaces in the table header field names and then performing a complete consistency comparison again; the third level performs fuzzy matching (303), determining whether the table header field names contain keywords from the standard field name library. The three-level matching strategy is executed sequentially according to priority. If any level matches successfully, the mapping success processing is executed (304), and the mapping result is displayed to the user for confirmation through the mapping result confirmation interface (307). If none of the three levels match, the mapping failure prompt is executed (305), requiring the user to manually specify the field correspondence. After the field mapping is completed, the system performs void and invalid data filtering for negative reviews (204). The void detection rules include: minimum effective character threshold judgment, evaluation text with less than 4 effective characters is judged as invalid evaluation and filtered; pure repetitive text recognition, the continuous repetitive character matching rule is used to detect text composed entirely of the same repeated character, and if the match is successful, it is judged as invalid evaluation and filtered. This system is deployed on Windows terminal devices, relying on local offline inference deployment components to build a local offline large language model inference environment, and using local embedded storage modules to store all evaluation data and judgment results. The entire processing flow runs offline, and the travel evaluation data of passengers and drivers are encrypted and stored locally, without transmitting any data to the external network. For example, import a negative review text as: The driver deliberately took a detour during the journey, and the interior of the carriage was dirty and smelly. After cleaning the text, the system retrieves the classification standards from the regulatory rule base, concatenates them into standardized prompt words, and then submits them to a local offline large language model for semantic analysis. The model output includes: risk level is medium risk, problem label is service quality issue, and the handling suggestion is to interview the driver and order rectification. Based on this, the system automatically generates a standard format pre-litigation handling work order and exports it as an editable handling work order file.

[0013] Example 2: Artificial Intelligence Judgment Process. This example details the workflow of the offline artificial intelligence judgment module (102). After the negative review data is imported and invalidated, the system enters the artificial intelligence semantic judgment stage. First, the rule configuration authorization module (106) loads the user-pre-configured ride-hailing regulatory rules (205). The regulatory rules include negative review rating classification standards, risk level classification rules for various types of issues, and handling suggestion templates. Users can customize the rule content according to the specific regulatory requirements of the local transportation authority. Second, the system performs dynamic prompt word construction (206). The filtered negative review text is concatenated with the regulatory rule content, and a complete model input prompt word is constructed according to the preset prompt word template format. The prompt word clearly specifies the content structure that the model needs to output, including three dimensions: risk level, issue type, and handling suggestion. Subsequently, the constructed prompt word is sent to the local offline large language model inference service (108). The local offline large language model inference service (108) runs large language model inference offline on the local device (207). The inference temperature parameter is set to 0.3 to ensure the stability and consistency of the output results. The maximum output characters for a single inference are limited to 1024 characters, and the inference timeout threshold is set to 120 seconds. If the inference process does not return a result after 120 seconds, the system triggers the judgment exception handling and retry (407) mechanism to automatically re-initiate the inference request. After the model returns the result, the system performs judgment result parsing (405) to extract the risk level, problem type, and disposal suggestion structured fields from the model output text. After parsing, the structured judgment result is stored in the local embedded storage module (107) to complete the structured judgment result storage (208) and ensure that all evaluation data and judgment results are persistently stored locally.

[0014] Example 3: Work Order Generation and Export. This example illustrates the automatic generation and export process of pre-emptive handling work orders. After the assessment results are stored in the local embedded storage module (107), the handling work order export unit (104) automatically generates standardized handling work orders (209) according to the assessment results in the database and the preset pre-emptive handling work order template. The work order content includes the following information: evaluation source information (passenger evaluation time, evaluation content, route information), driver information (name, professional qualification number, affiliated company), assessment conclusion (risk level, problem type, problem description), and handling suggestions (handling measures generated according to regulatory rules). After the work order is generated, the user can choose to export it individually or in batches. The handling work order export unit (104) writes the work order content into an editable document format file according to the standard format. The exported document is in an editable format, and the handling personnel can directly modify, supplement, and print it in the document. When exporting in batches, the system generates an independent document file for each work order. The file naming rules include the driver's name and the handling date information.

[0015] Example 4: Rule Configuration and Authorization Verification. This example illustrates the workflow of the rule configuration authorization module (106). The rule configuration authorization module (106) provides a custom configuration function for regulatory rules. Users can configure differential rating and classification standards according to the online ride-hailing service quality supervision regulations issued by the local transportation authorities. The rule configuration content includes: problem type definition (safety problems, service quality problems, price problems, attitude problems), risk level classification standards for each level, and corresponding handling suggestion templates for each problem type. When the system starts, the rule configuration authorization module (106) executes the authorization verification process. The authorization verification adopts a local encrypted authorization file verification method to verify the authorization period, authorized device identifier, and usage limit in the authorization file. After the authorization verification passes, the system enters normal working mode; if the authorization verification fails, the system restricts the use of the artificial intelligence analysis function and the work order export function. The rule configuration data is stored in the local embedded storage module (107), and the configuration modification takes effect in real time to ensure that the regulatory rules used in the subsequent artificial intelligence analysis process are consistent with the user's latest configuration.

[0016] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, substitutions, or improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligently analyzing and proactively handling negative reviews of ride-hailing services, comprising a process of importing review data, text processing, and generating a handling work order; characterized in that, Risk assessment is conducted based on a local offline large-scale model, specifically including the following steps: S1. Data Import and Smart Field Mapping Steps: Receive the externally imported ride-hailing negative review data file, parse and extract the multi-level table headers in the file, and use a multi-strategy matching method to establish a mapping relationship between the source data fields and the system standard fields to complete the automatic import and field alignment of the negative review data. S2, Filtering steps for empty and invalid negative review content: Perform void detection on the imported negative review text content, identify and filter invalid review data with empty content and no substantial information, and retain valid negative review records to enter the subsequent analysis process; S3, Custom Evaluation Rule Loading and Dynamic Prompt Word Construction Steps: Load the user-preconfigured evaluation rules and classification standards, dynamically construct structured prompt words in combination with the effective negative review text content, and encapsulate the negative review text, evaluation rules and judgment requirements into an inference request that conforms to the input format of the large language model; S4. Local offline large language model AI semantic judgment steps: Send the constructed inference request to the locally deployed offline large language model inference service, complete semantic understanding and risk level judgment in a local environment completely disconnected from the external network, and output a structured judgment result containing risk category, severity and handling suggestions. S5. Structured analysis results storage step: Write the structured analysis results into the local embedded storage module to realize the persistent storage of analysis data and subsequent retrieval and query; S6. Automatically generate standardized pre-complaint handling work orders: Based on the assessment results, automatically generate standardized pre-complaint handling work orders. The work order content includes driver information, negative review summary, risk assessment conclusion and suggested handling measures. S7. Steps for batch exporting work orders as editable documents: Batch export the generated work orders as editable files in standard office document format for business personnel to view, edit and print offline.

2. The method according to claim 1, characterized in that, In the S2 step of filtering empty and invalid negative review content, the following filtering rules are adopted: when the number of characters in the negative review text is less than or equal to 4 characters, it is determined to be an invalid review with empty content and is filtered out; at the same time, the continuous repeating character matching rule is adopted to identify invalid negative reviews that are spamming. When the entire negative review text consists of the same character repeated continuously, it is determined to be an invalid review that is spamming and is filtered out.

3. The method according to claim 1, characterized in that, In the data import and intelligent field mapping step S1, the multi-strategy matching method adopts a three-level priority matching strategy, specifically: the first level is exact matching, which compares the source data table header field name with the system standard field name one by one; the second level is whitespace-removed exact matching, which removes leading and trailing spaces from the source data table header field name and then compares it with the system standard field name again; the third level is field inclusion fuzzy matching, which determines whether the source data table header field name contains keyword fragments of the system standard field name, and executes them in sequence according to the three-level priority. When a higher-level match is successful, the lower-level match is no longer executed.

4. The method according to claim 1, characterized in that, In the local offline large language model AI semantic judgment step in S4, the following inference parameters are configured for the offline large language model inference service: the sampling temperature parameter is set to 0.3 to reduce the randomness of the model output and improve the stability and consistency of the judgment results; the maximum number of tokens generated in a single inference is set to 1024 to ensure the complete output of the structured judgment results; the inference timeout is set to 120 seconds, and when a single inference response exceeds this time threshold, it will be automatically interrupted and a retry mechanism will be triggered.

5. The method according to claim 1, characterized in that, The method further includes a device authorization verification step: when the system starts, a device machine code of the current device is generated by a hash algorithm, the authorization code is verified by a key-based hash message authentication code algorithm, and the authorization information is bound to the device machine code and stored in local encrypted storage; Each time the system starts, the device machine code is recalculated and compared with the stored authorization information. Only after the verification is successful can the data processing flow be entered; otherwise, the system will refuse to start.

6. The method according to claim 1, characterized in that, In the step of batch exporting editable documents from work orders in S7, an independent editable document file is generated for each work order. The naming rules for the document file include the driver's name and the date of disposal to enable quick identification and management of work order files. The exported document file adopts the standard office document format and supports editing, annotation, and printing operations in conventional office software.

7. A smart system for analyzing and handling negative reviews of ride-hailing services before complaints are filed, comprising data import, text processing, and work order output modules; characterized in that, Equipped with an offline AI analysis module, the system can operate independently in an intranet environment. The system includes: The data import mapping module (101) is used to receive externally imported ride-hailing negative review data files, parse and extract the multi-level table headers in the file, and use a multi-strategy matching method to establish a mapping relationship between the source data fields and the system standard fields, thereby completing the automatic import and field alignment of negative review data. The evaluation table data parsing unit (103) is set inside the data import mapping module (101) and is used to parse the imported evaluation data table format negative evaluation data file and extract the table header information and data row content. The offline artificial intelligence judgment module (102) is used to load the evaluation rules configured by the user, construct structured prompt words and send them to the local offline large language model reasoning service for semantic judgment, and output structured judgment results including risk category, severity and disposal suggestions; The document generation module (104) is used to automatically generate standardized pre-litigation handling work orders based on the assessment results, and to batch output the work orders as editable files in standard office document format; The local desktop background service module (105) is used to provide local background service support required for system operation and coordinate data interaction and task scheduling between modules. The rule configuration authorization module (106) is used to manage the configuration loading of evaluation rules and the verification of authorized devices used by the system; The local embedded storage module (107) is used to persistently store imported negative review data, analysis results, handling work orders and system configuration information; The local offline large language model inference service (108) provides large language model inference capabilities in a local environment that is completely disconnected from the external network, receives structured prompt word requests and returns semantic judgment results.

8. The system according to claim 7, characterized in that, In the data import mapping module (101), the multi-strategy matching method adopts a three-level priority matching strategy, specifically: the first level is exact matching, which compares the source data table header field name with the system standard field name one by one; the second level is whitespace-removing exact matching, which removes leading and trailing spaces from the source data table header field name and then compares it with the system standard field name again; the third level is field inclusion fuzzy matching, which determines whether the source data table header field name contains keyword fragments of the system standard field name; the data import mapping module (101) also includes a table header field extraction unit, which is used to automatically identify and extract field name information in the multi-layer merged table header structure of the evaluation data table file.

9. The system according to claim 7, characterized in that, In the offline artificial intelligence judgment module (102), the structured prompt words are constructed by concatenating and encapsulating the valid negative review text, the user-configured evaluation rules and classification standards, and the judgment output format requirements into a reasoning request that conforms to the input format of the large language model. When the offline artificial intelligence judgment module (102) calls the local offline large language model reasoning service (108), the reasoning parameters are configured as follows: sampling temperature parameter 0.3, maximum number of tokens generated in a single reasoning 1024, reasoning timeout time 120 seconds and support for automatic retry after timeout.

10. The system according to claim 7, characterized in that, In the rule configuration authorization module (106), the authorization verification process is as follows: when the system starts, the device machine code of the current device is generated by a hash algorithm, and the authorization code is verified by a key-based hash message authentication code algorithm. After verification, the authorization information is bound to the device machine code and written to local encrypted storage. Each time the system starts, the device machine code is recalculated and compared with the stored authorization information for verification. The rule configuration authorization module (106) is also responsible for managing the loading, editing and version control of user-defined evaluation rules.

Images