A malicious complaint identification method, device and equipment based on a large model and a medium

Abstract

The application discloses a malicious complaint identification method and device based on a large model, equipment and medium, relates to the technical field of artificial intelligence, and comprises the following steps: converting complaint voice into initial text data; performing word segmentation processing on the initial text data to obtain segmented text data, and performing vectorization processing on the segmented text data to obtain vector data; extracting target feature information from the segmented text data, determining a clustering cluster to which the segmented text data belongs based on each target vector data, and determining a clustering cluster label corresponding to the clustering cluster; identifying malicious complaints in the complaint voice based on the target feature information, the clustering cluster label, the target vector data and a malicious complaint identification large model; and the target malicious complaint identification large model is configured with an incremental learning strategy. Through the preset incremental learning strategy in the model, dynamic adaptation to emerging malicious complaint modes is realized, and the problem of insufficient generalization ability of traditional machine learning models is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a malicious complaint identification method and device based on a large model, equipment and a medium. BACKGROUND

[0002] The work hotline is an important channel for the public to express their demands. Its content is extensive and involves multiple functional departments, reflecting various issues of public concern. However, in the heavy hotline work, malicious complaints also need to be faced. Malicious complaints refer to the behavior of making demands on the staff beyond the reasonable scope or against the normal order through fabricating facts, repeated reporting, blackmailing and other means, in order to make profits, vent anger, harass or abuse rights, thereby occupying public resources and disrupting the normal order.

[0003] Currently, machine learning technology is usually used to identify malicious complaints of users. However, this method has the problem that the generalization ability of the machine learning model is limited, and when new malicious complaint methods or data distribution changes occur, the recognition effect of the model will decrease. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a malicious complaint identification method and device based on a large model, which can realize dynamic adaptation to emerging malicious complaint patterns through the pre-set incremental learning strategy in the model, and solve the problem of insufficient generalization ability of traditional machine learning models. The specific scheme is as follows:

[0005] In a first aspect, the present application provides a malicious complaint identification method based on a large model, comprising:

[0006] Real-time acquisition of complaint voice of a target user from a target voice call system, and conversion of each of the complaint voice into corresponding initial text data;

[0007] Performing word segmentation processing on each of the initial text data using a pre-set word segmentation tool to obtain corresponding segmented text data, and performing vectorization processing on each of the segmented text data to obtain corresponding target vector data;

[0008] Automatic extraction of target feature information from each of the segmented text data, determination of a clustering cluster to which each of the segmented text data belongs based on each of the target vector data and a target clustering algorithm, and determination of a clustering cluster label corresponding to each of the clustering clusters; wherein the target feature information includes a complaint occurrence time, an entity relationship and a user demand corresponding to each of the segmented text data, respectively;

[0009] input the target feature information, the clustering cluster label and the target vector data into a target malicious complaint identification large model, and identify malicious complaints in the complaint voice based on the target malicious complaint identification large model; wherein, an incremental learning strategy is pre-configured in the target malicious complaint identification large model.

[0010] Optionally, the converting each complaint voice into corresponding initial text data comprises:

[0011] characteristics of each complaint voice are obtained, and the speech sequence of each complaint voice is converted into a text sequence based on the target acoustic characteristics to generate the initial text data.

[0012] Optionally, before the inputting the target feature information, the clustering cluster label and the target vector data into the target malicious complaint identification large model, the method further comprises:

[0013] obtaining initial historical complaint data, and performing data format conversion on the initial historical complaint data to obtain corresponding target historical complaint data; wherein, the initial historical complaint data comprises data of several data types;

[0014] performing synonym replacement and sentence adjustment on the target historical complaint data by using natural language processing technology to obtain enhanced data, and generating virtual malicious complaint samples based on the enhanced data and target business rules;

[0015] obtaining an initial malicious complaint identification large model, and training the initial malicious complaint identification large model based on the enhanced data and the virtual malicious complaint samples to obtain a trained malicious complaint identification large model;

[0016] obtaining target feedback information, determining the target feedback information as a target reward signal, and adjusting the model parameters of the trained malicious complaint identification large model by using the target reward signal to obtain the target malicious complaint identification large model; wherein, the target feedback information includes the score of the model inference result of the trained malicious complaint identification large model by a target evaluation personnel.

[0017] Optionally, the identifying malicious complaints in the complaint voice based on the target malicious complaint identification large model comprises:

[0018] obtaining the malicious probability value corresponding to each segmented text data by using the target malicious complaint identification large model; wherein, the malicious probability value is a probability value representing that the complaint voice corresponding to the segmented text belongs to a malicious complaint;

[0019] If the malicious probability value corresponding to any segmented text data is greater than the preset probability threshold, the complaint voice corresponding to the any segmented text data is determined as a malicious complaint.

[0020] Optionally, the malicious complaint identification method based on a large model further includes:

[0021] behavior feature information of each target user corresponding to each segmented text data is determined, and each behavior feature information is quantitatively processed to obtain corresponding quantitative information; wherein the behavior feature information includes historical complaint records, complaint frequency and complaint change of each target user.

[0022] Each quantitative information is analyzed by using the target malicious complaint identification large model to determine malicious complaints in each complaint voice.

[0023] Optionally, after identifying malicious complaints in the complaint voice based on the target malicious complaint identification large model, the method further includes:

[0024] If any complaint voice is a malicious complaint, a malicious complaint type corresponding to the any complaint voice is determined, a preset warning strategy corresponding to the any complaint voice is determined from a preset template library based on the malicious complaint type, and a risk warning is performed on the any complaint voice according to the warning strategy.

[0025] Optionally, the malicious complaint identification method based on a large model further includes:

[0026] The target voice call system is monitored in real time, and if the number of new complaint voices in the target voice call system is not less than a preset number threshold, the model parameters of the target malicious complaint identification large model are fine-tuned by using an incremental learning strategy and each new complaint voice.

[0027] In a second aspect, the application provides a malicious complaint identification device based on a large model, comprising:

[0028] A voice conversion module is configured to acquire complaint voices of target users from a target voice call system in real time, and convert each complaint voice into corresponding initial text data.

[0029] A data vectorization module is configured to perform segmentation processing on each initial text data by using a preset segmentation tool to obtain corresponding segmented text data, and perform vectorization processing on each segmented text data to obtain corresponding target vector data.

[0030] The feature information extraction module is used to automatically extract target feature information from each of the segmented text data, determine the cluster to which each of the segmented text data belongs based on each of the target vector data and the target clustering algorithm, and determine the cluster label corresponding to each of the clusters; wherein, the target feature information includes the complaint occurrence time, entity relationship and user demand corresponding to each of the segmented text data respectively;

[0031] The malicious complaint identification module is used to input the target feature information, the cluster label and the target vector data into the target malicious complaint identification model, and to identify malicious complaints in the complaint voice based on the target malicious complaint identification model; wherein, the target malicious complaint identification model is pre-configured with an incremental learning strategy.

[0032] Thirdly, this application provides an electronic device, comprising:

[0033] Memory, used to store computer programs;

[0034] A processor for executing the computer program to implement the aforementioned large-model-based malicious complaint identification method.

[0035] Fourthly, this application provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the aforementioned malicious complaint identification method based on a large model.

[0036] This application first acquires the complaint voice of the target user in real time from the target voice call system, and converts each complaint voice into corresponding initial text data. Then, it uses a preset word segmentation tool to segment each initial text data to obtain corresponding segmented text data, and performs vectorization processing on each segmented text data to obtain corresponding target vector data. Next, it automatically extracts target feature information from each segmented text data, determines the cluster to which each segmented text data belongs based on the target vector data and a target clustering algorithm, and determines the cluster label corresponding to each cluster. The target feature information includes the complaint occurrence time, entity relationship, and user request corresponding to each segmented text data. Finally, the target feature information, the cluster label, and the target vector data are input into a large-scale target malicious complaint identification model, and the malicious complaint in the complaint voice is identified based on the large-scale target malicious complaint identification model. The large-scale target malicious complaint identification model is pre-configured with an incremental learning strategy. Therefore, this application achieves multi-dimensional structured analysis of text data through automatic word segmentation, vectorization, and feature extraction, overcoming the narrow coverage of traditional keyword matching. By generating cluster labels through clustering algorithms and combining target feature information with vector data into a large model, it realizes deep semantic and behavioral correlation analysis of complaint patterns, improving the recognition accuracy of complex malicious complaints. Through the incremental learning strategy pre-built in the model, it achieves dynamic adaptation to emerging malicious complaint patterns, solving the problem of insufficient generalization ability of traditional machine learning models. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 This application discloses a flowchart of a malicious complaint identification method based on a large model.

[0039] Figure 2 This is a schematic diagram of a specific malicious complaint identification method based on a large model disclosed in this application;

[0040] Figure 3 This is a flowchart of a model training method disclosed in this application;

[0041] Figure 4 This is a schematic diagram of a model training method disclosed in this application;

[0042] Figure 5 This is a schematic diagram of a malicious complaint identification device based on a large model disclosed in this application;

[0043] Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Current methods for identifying malicious complaints suffer from a decline in performance when faced with new malicious complaint tactics or changes in data distribution. To address this, this application provides a malicious complaint identification method based on a large model. Through a pre-defined incremental learning strategy, the method dynamically adapts to emerging malicious complaint patterns, thus resolving the issue of insufficient generalization ability in traditional machine learning models.

[0046] See Figure 1 As shown in the figure, this invention discloses a malicious complaint identification method based on a large model, including:

[0047] Step S11: Obtain the complaint voice of the target user in real time from the target voice call system, and convert each complaint voice into corresponding initial text data.

[0048] The malicious complaint identification method based on a large model in this embodiment includes two stages: the construction and application of a malicious complaint identification model for a work hotline. The specific process is as follows: Figure 2 As shown:

[0049] During the construction phase, multi-source data, including voice, text, user information, and business data, are collected and integrated from various sources. Natural language processing techniques are used for semantic augmentation and negative sample generation to achieve dynamic data enhancement. A sensitive word library is constructed by combining rule-based and text mining techniques, and a large-scale supervised learning model adapted to job characteristics is trained. Human feedback reinforcement learning and incremental learning strategies are also employed to ultimately train a large-scale malicious complaint identification model that integrates multi-dimensional features.

[0050] In the application stage, the real-time work orders are subjected to unified session modality, data word segmentation and vectorization processing, and work order elements such as time, location, entity and demands are extracted and the data is structured. After clustering similar work orders, the processed data is input into the large model, and dual recognition based on text features (semantic patterns such as sensitive word frequency, sentiment tendency, etc.) and user behavior features (historical complaint frequency, demand change, etc.) is performed to achieve high-precision malicious complaint determination. Finally, risk warnings for violations and violation information of malicious complaints are triggered, and corresponding handling suggestions are generated.

[0051] This embodiment solves the problems of low efficiency and narrow coverage dimension of the traditional method for malicious complaint recognition, and significantly improves the accuracy, timeliness and automation level of malicious complaint recognition on the work hotline.

[0052] In this embodiment, the process of converting each complaint voice into corresponding initial text data may specifically include:

[0053] Using the target speech recognition algorithm to perform feature analysis on each complaint voice to obtain the target acoustic features corresponding to each complaint voice respectively, and based on the target acoustic features, converting the speech sequence of each complaint voice into a text sequence to generate initial text data.

[0054] Specifically, configure a speech recognition engine based on the ASR (Automatic Speech Recognition) service, connect to the work hotline voice call system, and receive complaint voice call data (i.e., complaint voice) in real time. Use the deep neural network speech recognition algorithm (i.e., the target speech recognition algorithm) to extract acoustic features and match the language model for the speech data, convert the speech sequence into the corresponding text sequence, and generate a complaint text record (i.e., initial text data), ensuring the transcription accuracy and providing basic data for subsequent text analysis.

[0055] Step S12: Use a preset word segmentation tool to perform word segmentation processing on each of the initial text data to obtain the corresponding text data after word segmentation, and perform vectorization processing on each of the text data after word segmentation to obtain the corresponding target vector data. [[ID=X]]

[0056] In this embodiment, it is necessary to call a word segmentation tool to segment the complaint text according to the custom dictionary in the work field and the general Chinese word segmentation rules, and decompose the continuous text string into a sequence of lexical units with independent semantics. At the same time, identify and filter out the stop words in the text, such as words like "de", "le", "shi", etc. that are of little significance for malicious complaint recognition, and extract the core words to simplify the data dimension.

[0057] Then, a pre-trained word vector model or a self-supervised learning text vectorization method is used to map the segmented words (i.e., the segmented text data) into a high-dimensional vector space. Through semantic similarity calculation, words with similar meanings are placed close together in the vector space, transforming the semantic information of the text into computable numerical vectors (i.e., target vector data), providing quantified data for subsequent model processing.

[0058] Step S13: Automatically extract target feature information from each of the segmented text data, determine the cluster to which each of the segmented text data belongs based on each of the target vector data and the target clustering algorithm, and determine the cluster label corresponding to each of the clusters; wherein, the target feature information includes the complaint occurrence time, entity relationship and user demand corresponding to each of the segmented text data respectively.

[0059] In this embodiment, the process of automatically extracting target feature information (work order elements) from each segmented text data is as follows:

[0060] Time extraction: Using regular expression matching and named entity recognition technology, the date and time format strings in the complaint text are accurately located, the specific time information of the complaint event is extracted, and it is standardized into a unified time format of "YYYY-MM-DD HH:MM:SS", that is, the format of "year-month-day hour-minute-second", which facilitates the subsequent analysis of the timeliness of the complaint by time dimension.

[0061] Entity Extraction: Utilizing dependency parsing and named entity recognition algorithms, combined with entity type definitions (department name, staff position, service item name, etc.) from the work business knowledge graph, relevant entity objects are identified and extracted from the text. Simultaneously, the extracted entities are labeled and associated to construct an entity relationship network, providing support for tracing the subject and object of complaints.

[0062] Request Extraction: By employing text classification and summarization techniques, the core request paragraphs of the complaint text are located. Sentences expressing requests, desires, and demands are analyzed, and semantic role labeling is used to identify the sender, recipient, and specific content of the request. A concise statement of the request is generated and stored in a structured format for subsequent targeted evaluation of its reasonableness.

[0063] In addition, the process of clustering work orders in this embodiment is as follows:

[0064] Clustering Model Construction: A clustering algorithm suitable for text data was selected, and combined with the feature vectors after text vectorization, a cluster center initialization strategy, a distance metric, and a method for determining the number of clusters were set. The clustering model was trained on a historical work order dataset, enabling it to automatically divide the data into clusters based on the semantic similarity of the text.

[0065] Work order clustering: The text of the complaint work order to be processed is vectorized and input into the trained clustering model. The model automatically calculates the distance between the work order and the center of each cluster based on the semantic features of the text, and assigns the work order to the cluster with the most similar semantics. This realizes the classification and aggregation of different malicious complaint patterns or similar normal complaint scenarios, and improves the efficiency of subsequent identification.

[0066] Post-processing of clustering results: The quality of work orders in each cluster is assessed and analyzed. The density of work orders within a cluster and the separation between work orders between clusters are calculated, and abnormal clusters or misclassified work orders are removed. Simultaneously, representative feature words and semantic patterns of each cluster are extracted to form cluster labels, facilitating rapid retrieval and analysis of similar malicious complaint cases.

[0067] Step S14: Input the target feature information, the cluster label and the target vector data into the target malicious complaint identification model, and identify the malicious complaint in the complaint voice based on the target malicious complaint identification model; wherein, the target malicious complaint identification model is pre-configured with an incremental learning strategy.

[0068] In this embodiment, before inputting the target feature information, cluster labels and target vector data into the target malicious complaint identification model, the method further includes: obtaining initial historical complaint data and converting the initial historical complaint data into a data format to obtain the corresponding target historical complaint data.

[0069] Natural language processing (NLP) techniques are used to perform synonym replacement and sentence structure adjustment on historical complaint data to obtain enhanced data. Based on the enhanced data and target business rules, virtual malicious complaint samples are generated. An initial malicious complaint identification model is obtained, and the initial malicious complaint identification model is trained based on the enhanced data and virtual malicious complaint samples to obtain a trained malicious complaint identification model. Target feedback information is obtained, which is identified as the target reward signal. The model parameters of the trained malicious complaint identification model are adjusted using the target reward signal to obtain a target malicious complaint identification model.

[0070] In this embodiment, the process of identifying malicious complaints in complaint voice based on the target malicious complaint identification big model may specifically include: using the target malicious complaint identification big model to obtain the malicious probability value corresponding to each segmented text data; wherein, the malicious probability value is the probability value that the complaint voice corresponding to the segmented text belongs to a malicious complaint; if the malicious probability value corresponding to any segmented text data is greater than a preset probability threshold, then the complaint voice corresponding to any segmented text data is determined to be a malicious complaint.

[0071] In addition, the malicious complaint identification method based on a large model in this embodiment further includes: determining the behavioral feature information of the target user corresponding to each segmented text data, quantifying each behavioral feature information to obtain corresponding quantitative information; wherein, the behavioral feature information includes the historical complaint records, complaint frequency and change of request corresponding to each target user; and using the target malicious complaint identification large model to analyze each quantitative information to determine the malicious complaint in each complaint voice.

[0072] In other words, the malicious complaint identification method in this embodiment can be divided into two types:

[0073] Malicious Complaint Identification Based on Text Features: The preprocessed complaint text feature vector is input into a large model. The model, based on semantic pattern differences learned from massive amounts of malicious and legitimate complaint text data (frequency of sensitive words, logical coherence of sentences, text sentiment, etc.), calculates the probability value through nonlinear transformation and feature combination using a multi-layer neural network. If the probability value exceeds a set threshold, it is determined to be a malicious complaint.

[0074] Malicious Complaint Identification Based on Behavioral Characteristics: This method integrates behavioral characteristic data such as complainant's historical complaint records, complaint frequency, distribution of complaint targets, and changes in requests, quantifies these data, and inputs them into a large model. The model combines the results of text feature analysis to comprehensively assess whether the complainant's behavioral patterns indicate malicious intent (such as frequently making unreasonable complaints against the same department or frequently changing requests to achieve improper purposes), further improving the accuracy and comprehensiveness of malicious complaint identification.

[0075] It should be noted that this embodiment can combine the two malicious complaint identification methods according to actual needs. In one specific implementation, the large model analyzes text features based on text vectors to calculate a first probability value; the text features include the frequency of sensitive words, sentiment tendency, and logical coherence; the large model analyzes behavioral patterns based on the user behavior features to calculate a second probability value; the behavioral patterns include the frequency of historical complaints, the concentration of complaint targets, and the number of times the request has changed; the first probability value and the second probability value are weighted and fused to obtain a comprehensive score; if the comprehensive score exceeds a preset probability threshold, it is determined to be a malicious complaint.

[0076] By analyzing multiple dimensions, including frequency statistics of sensitive words in the text, contextual sentiment analysis, and semantic logic verification, we can deeply uncover clues to malicious intent within complaint texts. This comprehensive approach improves accuracy in determining malicious intent. By integrating dynamic behavioral characteristics such as users' historical complaint frequency, changing complaint patterns, and consistency of complaint content, we can construct precise user profiles, effectively identifying habitual or intentionally malicious complainants.

[0077] By automatically analyzing the semantics of work order texts and user behavior characteristics using a large-scale model, malicious complaints can be quickly identified, solving the pain point of time-consuming traditional manual screening. This improves the efficiency of identifying and blocking malicious complaints by several orders of magnitude. Furthermore, it achieves full-process intelligent work order processing, significantly shortening the work order cycle and freeing up human review resources to focus on more complex inquiries and services, thus optimizing the hotline's human resource allocation.

[0078] In this embodiment, after identifying malicious complaints in the complaint voice based on the target malicious complaint identification model, the method further includes: if any complaint voice is a malicious complaint, determining the malicious complaint type corresponding to any complaint voice, determining the warning strategy corresponding to any complaint voice from a preset template library based on the malicious complaint type, and issuing a risk warning for any complaint voice according to the warning strategy. The specific process is as follows:

[0079] Violation Risk Warning: Based on the rule knowledge base and the mapping relationship of violation risk characteristics in malicious complaint cases, when the large model identifies a malicious complaint and the content of the complaint involves risks such as potentially infringing on the rights of the complainant or disrupting normal work order, a risk warning mechanism is triggered. Warning information is generated, detailing the potential risk points, the severity assessment of the risk, and the relevant regulatory basis, promptly alerting the hotline management personnel and legal department to intervene and handle the matter.

[0080] Warning of Violations: Combining a violation dissemination model with a knowledge base of violations from the work hotline, analyze the sensitivity, dissemination potential, and potential social reactions of malicious complaints. If a malicious complaint is determined to pose a risk of spreading negative violations, such as involving hot work topics or easily inciting adverse situations, a warning of violations will be activated. The warning information should include the risk level of the violation, predictions of possible dissemination channels and audience scope, and analysis of the evolution trend of the violation, to assist managers in formulating strategies to deal with violations.

[0081] Handling suggestion generation: Based on the malicious complaint identification results and corresponding warning types, a pre-set handling suggestion template library is invoked. For malicious complaints in different scenarios (such as maliciously framing staff, maliciously interfering with work service processes, etc.), and combined with experience from past successful handling cases, handling suggestion plans are generated, including specific information such as communication scripts with complainants, key points for evidence collection, internal coordination processes, and key points for external responses, to guide hotline staff in carrying out subsequent efficient and compliant handling work.

[0082] Therefore, this application achieves multi-dimensional structured analysis of text data through automatic word segmentation, vectorization, and feature extraction, overcoming the narrow coverage of traditional keyword matching. By generating cluster labels through clustering algorithms and combining target feature information with vector data into a large model, it realizes deep semantic and behavioral correlation analysis of complaint patterns, improving the recognition accuracy of complex malicious complaints. Through the incremental learning strategy pre-built in the model, it achieves dynamic adaptation to emerging malicious complaint patterns, solving the problem of insufficient generalization ability of traditional machine learning models.

[0083] As can be seen from the foregoing embodiments, this application describes the overall process of identifying malicious complaints using a large-scale malicious complaint identification model. To make the technical solution in this application more complete, the process of constructing the large-scale malicious complaint identification model will be described in detail below. See also Figure 3 As shown in the figure, an embodiment of the present invention discloses a method for training a large model for identifying malicious complaints, including:

[0084] Step S21: Obtain initial historical complaint data and convert the initial historical complaint data into a data format to obtain the corresponding target historical complaint data; wherein, the initial historical complaint data includes data of several data types.

[0085] The model building process in this embodiment is as follows: Figure 4 As shown, the process includes steps such as heterogeneous data acquisition, dynamic data enhancement, establishment of a sensitive word database, large model training, feedback learning, and incremental learning.

[0086] In this embodiment, a diversified data acquisition module is first constructed to connect to the work hotline system and receive heterogeneous data (i.e., initial historical complaint data), including voice recordings, transcribed text, basic information of complainants (such as age, gender, and past complaint records), business data of relevant departments related to the complaint incident (such as processing time and completion status), and corresponding timestamps. Using data extraction, transformation, and loading technologies, the collected data is uniformly converted into a preset data format (i.e., target historical complaint data) and stored in a data warehouse, laying the data foundation for subsequent processing and ensuring data integrity and accuracy.

[0087] Step S22: Use natural language processing technology to perform synonym replacement and sentence structure adjustment on the target historical complaint data to obtain corresponding enhanced data, and generate virtual malicious complaint samples based on the enhanced data and target business rules.

[0088] In this embodiment, a dynamic data augmentation mechanism is established based on historical complaint data. Semantic augmentation techniques from natural language processing algorithms are used to replace synonyms and adjust sentence structure in the complaint text, generating diverse textual samples without altering the original meaning. Simultaneously, in conjunction with business rules, different complaint scenarios are simulated to generate corresponding false complaint data as negative samples. The data augmentation strategy is regularly adjusted according to the dynamic changes in the hotline business, ensuring the dataset continuously reflects the actual business situation and improving the richness and representativeness of the data.

[0089] Furthermore, this embodiment, through the collaboration of domain experts and linguistics experts, and based on relevant regulations, disciplinary norms, and past cases of malicious complaints, has compiled a list of sensitive words with malicious intent. This list includes insulting terms, common expressions used in false accusations, and words implying unreasonable demands, thus initially constructing a basic sensitive word database. Text mining algorithms are used to analyze a large number of work-related complaint texts to discover potential emerging words with malicious associations. These words are then manually reviewed and added to the sensitive word database. The sensitive word database is regularly updated and maintained to ensure it aligns with the actual business environment and sensitively captures the linguistic features of malicious complaints.

[0090] Based on the linkage between a sensitive word database and a pre-set rule knowledge base, high-risk complaints that may disrupt public order can be identified in real time, and suggested solutions can be generated. Through clustering of similar work orders and analysis of hotspot maps, potential group complaints can be detected in advance.

[0091] Step S23: Obtain the initial malicious complaint identification model, and train the initial malicious complaint identification model based on the enhanced data and the virtual malicious complaint samples to obtain the trained malicious complaint identification model.

[0092] This embodiment selects a pre-trained language model (i.e., an initial large-scale model for identifying malicious complaints) as the basic architecture. Its input layer is adapted to the features of the hotline complaint data, and the embedding layer transforms complaint text, user information, and other data into vector representations. Using the dynamically augmented dataset, supervised learning is employed to train the model. A classification loss function is set, and model parameters are optimized, enabling the model to learn the different feature representations of malicious and normal complaints in terms of text semantics and user behavior patterns. During training, a cross-validation mechanism is introduced to evaluate the model's generalization ability, prevent overfitting, and ensure that the model has good recognition performance when faced with new data.

[0093] Step S24: Obtain target feedback information, determine the target feedback information as a target reward signal, and use the target reward signal to adjust the model parameters of the trained malicious complaint identification model to obtain the target malicious complaint identification model; wherein, the target feedback information includes the score of the target evaluator on the model inference result of the trained malicious complaint identification model.

[0094] This example requires building a human feedback collection platform. Hotline staff and professional inspectors are invited as evaluators to manually review cases initially determined by the model to be malicious or legitimate complaints, providing correct classification labels and evaluating the reasonableness of the model's judgment. Human feedback information (i.e., target feedback information) is integrated into reward signals and fed back into the training process of the large model. This adjusts the model's decision boundaries and weight parameters, enhancing the accuracy and reliability of the model's malicious complaint identification, and making the model's judgments more closely aligned with human cognition and judgment standards in real-world scenarios.

[0095] After training and optimization through the above stages, the model is considered complete when its performance metrics, such as accuracy and recall, for identifying malicious complaints on the validation set reach preset thresholds, and its operational stability and response speed in the actual hotline test environment meet business requirements. It is then deployed to the hotline system to automate the identification of malicious complaints from real-time complaint data, assisting staff in efficiently handling hotline matters and improving the quality and efficiency of hotline services.

[0096] In addition, it should be noted that this implementation can also monitor the target voice call system in real time. If the number of newly added complaint voices in the target voice call system is not less than the preset threshold, the model parameters of the target malicious complaint identification model are fine-tuned using an incremental learning strategy and each newly added complaint voice.

[0097] In other words, this embodiment can establish an incremental learning mechanism to continuously monitor the updates of hotline business data. When a certain amount of new complaint data is added, the old and new data are merged to incrementally train the model. During the training process, the model is fine-tuned using the new data to update its understanding of newly emerging malicious complaint patterns and expressions, while retaining the original model's ability to recognize known patterns. This ensures that the model can adapt to the development and changes in hotline business in a timely manner, continuously improving the timeliness and accuracy of malicious complaint identification.

[0098] By incrementally learning and updating emerging malicious complaint patterns (such as new sensitive words and variations in wording), the system can continuously absorb annotated new work order data, automatically capturing and updating the evolving characteristics of malicious complaint patterns, such as new sensitive words and modified wording techniques, giving the model dynamic adaptability. Combined with feedback from domain experts, the system continuously fine-tunes the model's decision boundaries, optimizing judgment thresholds and rule weights. This allows the system to continuously learn from the experience and judgment of human experts, achieving self-correction and intelligent evolution, thus improving long-term robustness and adaptability.

[0099] See Figure 5 As shown in the figure, an embodiment of the present invention discloses a malicious complaint identification device based on a large model, comprising:

[0100] The voice conversion module 11 is used to acquire the complaint voice of the target user from the target voice call system in real time and convert each complaint voice into corresponding initial text data;

[0101] The data vectorization module 12 is used to perform word segmentation processing on each of the initial text data using a preset word segmentation tool to obtain the corresponding word segmented text data, and to perform vectorization processing on each of the word segmented text data to obtain the corresponding target vector data.

[0102] The feature information extraction module 13 is used to automatically extract target feature information from each of the segmented text data, determine the cluster to which each of the segmented text data belongs based on each of the target vector data and the target clustering algorithm, and determine the cluster label corresponding to each of the clusters; wherein, the target feature information includes the complaint occurrence time, entity relationship and user demand corresponding to each of the segmented text data respectively;

[0103] The malicious complaint identification module 14 is used to input the target feature information, the cluster label and the target vector data into the target malicious complaint identification model, and to identify malicious complaints in the complaint voice based on the target malicious complaint identification model; wherein, the target malicious complaint identification model is pre-configured with an incremental learning strategy.

[0104] In some specific embodiments, the speech conversion module 11 may specifically include:

[0105] The speech conversion unit is used to perform feature analysis on each of the complaint voices using a target speech recognition algorithm to obtain the target acoustic features corresponding to each of the complaint voices, and to convert the speech sequence of each of the complaint voices into a text sequence based on the target acoustic features to generate the initial text data.

[0106] In some specific embodiments, the malicious complaint identification module 14 further includes:

[0107] A data format conversion unit is used to acquire initial historical complaint data and convert the initial historical complaint data into a data format to obtain corresponding target historical complaint data; wherein, the initial historical complaint data includes data of several data types;

[0108] The virtual sample generation unit is used to perform synonym replacement and sentence structure adjustment on the target historical complaint data using natural language processing technology to obtain corresponding enhanced data, and to generate virtual malicious complaint samples based on the enhanced data and target business rules.

[0109] The model training unit is used to obtain an initial malicious complaint identification model and train the initial malicious complaint identification model based on the enhanced data and the virtual malicious complaint samples to obtain a trained malicious complaint identification model.

[0110] A parameter adjustment unit is used to acquire target feedback information, determine the target feedback information as a target reward signal, and adjust the model parameters of the trained malicious complaint identification model using the target reward signal to obtain the target malicious complaint identification model; wherein, the target feedback information includes the score given by the target evaluator to the model inference result of the trained malicious complaint identification model.

[0111] In some specific embodiments, the malicious complaint identification module 14 may specifically include:

[0112] The probability value acquisition unit is used to acquire the malicious probability value corresponding to each of the segmented text data using the target malicious complaint identification big model; wherein, the malicious probability value is the probability value representing the probability that the complaint voice corresponding to the segmented text belongs to a malicious complaint;

[0113] The malicious complaint determination unit is used to determine the complaint voice corresponding to any segmented text data as a malicious complaint if the malicious probability value corresponding to any segmented text data is greater than a preset probability threshold.

[0114] In some specific embodiments, the malicious complaint identification device based on a large model further includes:

[0115] The feature information determination module is used to determine the behavioral feature information of the target user corresponding to each of the segmented text data, and to quantify each of the behavioral feature information to obtain corresponding quantitative information; wherein, the behavioral feature information includes the historical complaint records, complaint frequency and request change status corresponding to each of the target users.

[0116] The malicious complaint identification module is used to analyze the quantitative information using the target malicious complaint identification big data model to identify malicious complaints in each of the complaint voices.

[0117] In some specific embodiments, the malicious complaint identification module 14 further includes:

[0118] The risk warning unit is used to determine the type of malicious complaint corresponding to any complaint voice if any complaint voice is a malicious complaint, determine the warning strategy corresponding to any complaint voice from a preset template library based on the malicious complaint type, and issue a risk warning to any complaint voice according to the warning strategy.

[0119] In some specific embodiments, the malicious complaint identification device based on a large model further includes:

[0120] The parameter fine-tuning module is used to monitor the target voice call system in real time. If the number of newly added complaint voices in the target voice call system is not less than a preset threshold, the model parameters of the target malicious complaint identification model are fine-tuned using an incremental learning strategy and each of the newly added complaint voices.

[0121] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0122] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the malicious complaint identification method based on a large model disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be a computer.

[0123] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0124] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0125] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the large-model-based malicious complaint identification method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0126] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned malicious complaint identification method based on a large model. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0127] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0128] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0129] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0130] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for identifying malicious complaints based on a large model, characterized in that, include: The complaint voices of target users are acquired in real time from the target voice call system, and each complaint voice is converted into corresponding initial text data. The initial text data is segmented using a preset word segmentation tool to obtain the corresponding segmented text data, and the segmented text data is vectorized to obtain the corresponding target vector data. Target feature information is automatically extracted from each of the segmented text data. Based on each of the target vector data and the target clustering algorithm, the cluster to which each of the segmented text data belongs is determined, and the cluster label corresponding to each cluster is determined. The target feature information includes the complaint occurrence time, entity relationship, and user demand corresponding to each of the segmented text data. The target feature information, the cluster label, and the target vector data are input into the target malicious complaint identification model, and malicious complaints in the complaint voice are identified based on the target malicious complaint identification model; wherein, the target malicious complaint identification model is pre-configured with an incremental learning strategy.

2. The malicious complaint identification method based on a large model according to claim 1, characterized in that, The process of converting each of the complaint voice messages into corresponding initial text data includes: The target speech recognition algorithm is used to perform feature analysis on each of the complaint speech to obtain the target acoustic features corresponding to each of the complaint speech, and the speech sequence of each complaint speech is converted into a text sequence based on the target acoustic features to generate the initial text data.

3. The malicious complaint identification method based on a large model according to claim 1, characterized in that, Before inputting the target feature information, the cluster label, and the target vector data into the target malicious complaint identification model, the method further includes: Acquire initial historical complaint data and convert the initial historical complaint data into a data format to obtain corresponding target historical complaint data; wherein, the initial historical complaint data includes data of several data types; Natural language processing technology is used to perform synonym replacement and sentence structure adjustment on the target historical complaint data to obtain corresponding enhanced data, and virtual malicious complaint samples are generated based on the enhanced data and target business rules; An initial malicious complaint identification model is obtained, and the initial malicious complaint identification model is trained based on the enhanced data and the virtual malicious complaint samples to obtain a trained malicious complaint identification model. Obtain target feedback information, determine the target feedback information as a target reward signal, and use the target reward signal to adjust the model parameters of the trained malicious complaint identification model to obtain the target malicious complaint identification model; wherein, the target feedback information includes the score of the target evaluator on the model inference result of the trained malicious complaint identification model.

4. The malicious complaint identification method based on a large model according to claim 1, characterized in that, The process of identifying malicious complaints in the complaint speech based on the target malicious complaint identification big data model includes: The malicious complaint identification model is used to obtain the malicious probability value corresponding to each segmented text data; wherein, the malicious probability value is the probability value that the complaint voice corresponding to the segmented text belongs to a malicious complaint; If the probability value of malicious activity corresponding to any segmented text data is greater than a preset probability threshold, then the complaint voice corresponding to any segmented text data will be identified as a malicious complaint.

5. The malicious complaint identification method based on a large model according to claim 1, characterized in that, Also includes: The behavioral feature information of the target user corresponding to each segmented text data is determined, and the behavioral feature information is quantified to obtain corresponding quantitative information; wherein, the behavioral feature information includes the historical complaint records, complaint frequency and request change status corresponding to each target user; The target malicious complaint identification big data model is used to analyze the quantitative information to determine the malicious complaints in each of the complaint voices.

6. The malicious complaint identification method based on a large model according to claim 1, characterized in that, After identifying malicious complaints in the complaint speech based on the target malicious complaint identification big data model, the method further includes: If any complaint voice is a malicious complaint, then the type of malicious complaint corresponding to the complaint voice is determined, and based on the type of malicious complaint, the warning strategy corresponding to the complaint voice is determined from the preset template library, and a risk warning is given to the complaint voice according to the warning strategy.

7. The malicious complaint identification method based on a large model according to any one of claims 1 to 6, characterized in that, Also includes: The target voice call system is monitored in real time. If the number of newly added complaint voice messages in the target voice call system is not less than a preset threshold, the model parameters of the target malicious complaint identification model are fine-tuned using an incremental learning strategy and each of the newly added complaint voice messages.

8. A malicious complaint identification device based on a large model, characterized in that, include: The voice conversion module is used to acquire the complaint voice of the target user from the target voice call system in real time and convert each complaint voice into corresponding initial text data. The data vectorization module is used to perform word segmentation on each of the initial text data using a preset word segmentation tool to obtain the corresponding segmented text data, and to perform vectorization on each of the segmented text data to obtain the corresponding target vector data. The feature information extraction module is used to automatically extract target feature information from each of the segmented text data, determine the cluster to which each of the segmented text data belongs based on each of the target vector data and the target clustering algorithm, and determine the cluster label corresponding to each of the clusters; wherein, the target feature information includes the complaint occurrence time, entity relationship and user demand corresponding to each of the segmented text data respectively; The malicious complaint identification module is used to input the target feature information, the cluster label and the target vector data into the target malicious complaint identification model, and to identify malicious complaints in the complaint voice based on the target malicious complaint identification model; wherein, the target malicious complaint identification model is pre-configured with an incremental learning strategy.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the malicious complaint identification method based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the malicious complaint identification method based on a large model as described in any one of claims 1 to 7.

Images