Method for realizing automatic monitoring and anti-fraud of negative public opinions based on large model
By adopting the large-model federal learning method in the public opinion monitoring system, a large-scale user portrait model is solved, and the problems of narrow application scope of public opinion monitoring systems and low accuracy of user portrait feature models in the existing technology are solved, and the accurate identification of various public opinion events and user portrait features are achieved, reducing costs and workload.
Patent Information
- Application Number
- CN202411911404.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-24
- Publication Date
- 2025-04-25
AI Technical Summary
The existing public opinion monitoring system has a narrow scope of application and cannot effectively identify and deal with various types of public opinion events. The user portrait feature model lacks self-reasoning and generalization capabilities, resulting in low matching accuracy.
A federated learning method based on big models is adopted to build a large model for public opinion monitoring of user portraits, and distributed training is carried out through aggregation server and multiple data participants, and a large amount of high-quality private domain public opinion data are used for training, so as to achieve accurate identification of public opinion events and accurate identification of user portrait features.
It greatly improves the accuracy of public opinion event recognition, has the ability to accurately identify user portrait characteristics of various public opinion event makers, solves the contradiction between model training and data islands, and reduces the cost of building and maintaining models.
Smart Images
Figure CN119887234A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of public opinion monitoring, and in particular to a method for automatically monitoring negative public opinion and preventing fraud based on a large model. Background Art
[0002] Public opinion monitoring can analyze public opinion trends and help users better respond to public opinion events, thereby reducing the risk of being misled or deceived. Public opinion monitoring helps to implement user profiling and anti-fraud.
[0003] However, the existing public opinion monitoring systems commonly use technologies such as text data analysis, natural language processing, and machine learning, which all require manual collection of similar sample data and construction of exclusive models before they can accurately process public opinion data with similar characteristics. This approach has a narrow scope of application and can only be effective in specific scenarios or for public opinions with similar characteristics. The usual solution to the narrow scope of application is to manually collect more public opinion data with different characteristics, establish multiple public opinion models, or establish a public opinion model that is compatible with multiple features. The disadvantage of this solution is that the number of models built is large, the cost is high, and it is impossible to exhaust all types of models. There are also many disadvantages in building a public opinion monitoring model that basically takes into account multiple events. Often, such models have low recognition accuracy and are not sensitive enough.
[0004] The models used in existing technologies for building user portraits lack self-reasoning and generalization capabilities. The underlying principle is to build a feature model through historical data for matching. This feature model contains very limited information, resulting in low matching accuracy. To solve the problem of low accuracy of user portrait feature models, the general approach is to build N feature models, then match user information with the N feature models one by one, and finally use a certain strategy for comprehensive evaluation to improve accuracy, which is labor-intensive and costly. Summary of the invention
[0005] The purpose of the present invention is to address the problem of narrow application scope of public opinion monitoring systems in the background technology, and to propose a method for automatically monitoring negative public opinion and anti-fraud based on a large model.
[0006] The technical solution of the present invention is a method for automatically monitoring negative public opinion and preventing fraud based on a large model, comprising the following steps:
[0007] S1. Build a large model of user portraits for public opinion monitoring. This model is a federated model structure. The federated model is divided into two layers: the upper layer is the aggregation server, and the lower layer is multiple data participants.
[0008] S2. Training a large user portrait model for public opinion monitoring;
[0009] S3. Conduct public opinion monitoring and anti-fraud by training a large public opinion monitoring user portrait model.
[0010] Preferably, in S1, the aggregation server sends the global parameters to the data participants, and the data participants upload the model parameters and Fisher information to the aggregation server.
[0011] Preferably, data participants use global parameters to perform model training in a private domain, and the training data of data participants are not shared.
[0012] Preferably, data participants use global parameters to perform model training in both private and public domains, and private domain data is not shared, but public domain data is shared.
[0013] Preferably, in S2, training the public opinion monitoring user portrait big model includes the following steps: S201, the aggregation server initializes the public opinion monitoring user portrait big model parameters and sends them to each data participant; S202, each data participant collects private domain user public opinion corpus; S203, the participating units clean and organize the corpus into structured document data that meets the training requirements of the public opinion monitoring user portrait big model; S204, the data participants randomly extract a certain amount of training data for annotation as model training samples; S205, the data participants use the Transform architecture to train local sample data within the unit; S206, the data participants obtain local big model parameters and calculate Fisher information; S207, the data participants upload the big model parameters and Fisher information to the aggregation server; S208, the aggregation server performs federal resolution of global Fisher information; S209, the aggregation server integrates the local big model parameters uploaded by each data participant into global big model parameters according to the Fisher information; S210, generates a public opinion monitoring user portrait big model.
[0014] Preferably, in S3, public opinion monitoring and anti-fraud include the following steps: S301, real-time collection of public opinion corpus of complaints and comments from Internet or internal company users; S302, cleaning and organizing the corpus into structured document data that meets the training requirements of the public opinion monitoring user portrait big model; S303, inputting the public opinion monitoring user portrait big model for processing; S304, judging whether the current comment constitutes a public opinion event, if not, returning to step S301 to continue the loop; if yes, jumping to the next step to continue execution; S305, judging whether the public opinion event contains fraudulent nature; S306, if the public opinion information is just a hot event and will not cause property and economic losses to users, then the public relations department will pay attention to it simultaneously; S307, if the public opinion information will cause property and economic losses to users, then fix the evidence and synchronize with the legal department.
[0015] Compared with the prior art, the present invention has the following beneficial technical effects: the large model uses a large amount of high-quality private domain public opinion data for training, so the accuracy of public opinion event recognition is greatly improved, and it also has the ability to accurately identify the user portrait features of various public opinion event makers. Because the large model is trained in a distributed manner using a federated learning method, this method solves the contradiction between model training and data silos. Through the public opinion monitoring user portrait large model, a large model is implemented to monitor and identify all types of public opinion events, and a large model is used to replace N user portrait models to identify whether the public opinion event maker is involved in fraud. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A schematic diagram of the federation model proposed for the present invention;
[0017] Figure 2 A flowchart for training a large user portrait model for public opinion monitoring;
[0018] Figure 3 This is a flowchart of public opinion monitoring. DETAILED DESCRIPTION
[0019] Embodiment 1, as Figure 1-Figure 3 As shown, the present invention proposes a method for automatically monitoring negative public opinion and preventing fraud based on a large model, comprising the following steps:
[0020] S1. Build a large model of user portraits for public opinion monitoring. This model is a federated model structure. The federated model is divided into two layers: the upper layer is the aggregation server, and the lower layer is multiple data participants.
[0021] S2. Training a large user portrait model for public opinion monitoring;
[0022] S3. Conduct public opinion monitoring and anti-fraud by training a large public opinion monitoring user portrait model.
[0023] Embodiment 2: The present invention proposes a method for automatically monitoring negative public opinions and combating fraud based on a large model. Compared with embodiment 1, this embodiment introduces S1 in detail.
[0024] In S1, the aggregation server sends global parameters to data participants, and data participants upload model parameters and Fisher information to the aggregation server. Data participants use global parameters to train models in private domains, and their training data is not shared, or they train models in public and private domains, and private domain data is not shared, but public domain data is shared.
[0025] Embodiment three, as Figure 2 As shown, the present invention proposes a method for automatically monitoring negative public opinion and combating fraud based on a large model. Compared with the first embodiment, this embodiment introduces S2 in detail.
[0026] Training the public opinion monitoring user portrait big model includes the following steps: S201, the company's aggregation server initializes the parameters of the public opinion monitoring user portrait big model and sends it to each data participant. The data participants are participating units such as the company's internal, collaborative units, and industrial alliances; S202, each data participant, i.e., participating unit, collects user public opinion corpus from private domain customer service centers, forums, forums, communities, etc. In addition to its own private domain public opinion corpus, the company also collects public domain public opinion corpus simultaneously; S203, the participating units clean and organize the corpus into structured document data that meets the training requirements of the public opinion monitoring user portrait big model; S204, the data participants, i.e., participating units, randomly select a certain A large amount of training data is labeled as model training samples; S205, data participants, i.e. participating units, use the Transform architecture to train local sample data within the unit; S206, data participants, i.e. participating units, obtain local large model parameters and calculate Fisher information; S207, data participants, i.e. participating units, upload large model parameters and Fisher information to the aggregation server; S208, the aggregation server performs federal resolution of global Fisher information; S209, the aggregation server integrates the local large model parameters uploaded by each data participant, i.e. participating unit, into global large model parameters according to the Fisher information; S210, generates a large model of user portraits for public opinion monitoring.
[0027] Embodiment 4, as shown in 3, the present invention proposes a method for automatically monitoring negative public opinion and anti-fraud based on a large model. Compared with Embodiment 1, Embodiment 2 or Embodiment 3, this embodiment introduces S3 in detail.
[0028] Public opinion monitoring and anti-fraud include the following steps: S301, real-time or scheduled collection of public opinion corpus of user complaints and comments on the Internet or the company's internal customer service center, bulletin board, forum, community, etc.; S302, cleaning and organizing the corpus into structured document data that meets the training requirements of the public opinion monitoring user portrait large model; S303, inputting the public opinion monitoring user portrait large model for processing; S304, judging whether the current comment constitutes a public opinion event, if not, returning to step S301 to continue the loop; if yes, jumping to the next step to continue execution; S305, judging whether the public opinion event contains fraudulent nature, such as hyping up a surge or plunge in housing prices in a certain place, a sharp rise or fall in a certain stock, and other false information that misleads users; S306, if the public opinion information is just a hot event and will not cause economic losses such as property to users, then the public relations department will pay attention to it simultaneously; S307, if the public opinion information will cause economic losses such as property to users, then fix the evidence and synchronize with the legal department.
[0029] In summary, the big model uses a large amount of high-quality private domain public opinion data for training, so the accuracy of public opinion event recognition is greatly improved, and it also has the ability to accurately identify the user portrait features of various public opinion event makers. Because the federated learning method is used for distributed training of the big model, this method solves the contradiction between model training and data silos. Through the public opinion monitoring user portrait big model, a big model is implemented to monitor and identify all types of public opinion events, and a big model is used to replace N user portrait models to identify whether the public opinion event maker is involved in fraud.
[0030] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.
Claims
1. A method for automatically monitoring negative public opinion and anti-fraud based on a large model, characterized in that: The following steps are involved: S1. Build a large model of user portraits for public opinion monitoring. This model is a federated model structure. The federated model is divided into two layers: the upper layer is the aggregation server, and the lower layer is multiple data participants. S2. Training a large user portrait model for public opinion monitoring; S3. Conduct public opinion monitoring and anti-fraud by training a large public opinion monitoring user portrait model.
2. According to claim 1, the method for automatically monitoring negative public opinion and preventing fraud based on a large model is characterized in that: In S1, the aggregation server sends the global parameters to the data participants, and the data participants upload the model parameters and Fisher information to the aggregation server.
3. According to claim 2, the method for automatically monitoring negative public opinion and preventing fraud based on a large model is characterized in that: Data participants use global parameters to train models in private domains, and their training data is not shared.
4. According to claim 2, the method for automatically monitoring negative public opinion and preventing fraud based on a large model is characterized in that: Data participants use global parameters to train models in private and public domains. Private domain data is not shared, but public domain data is shared.
5. According to claim 1, the method for automatically monitoring negative public opinion and preventing fraud based on a large model is characterized in that: In S2, training the public opinion monitoring user portrait big model includes the following steps: S201, the aggregation server initializes the public opinion monitoring user portrait big model parameters and sends them to each data participant; S202, each data participant collects private domain user public opinion corpus; S203, the participating units clean and organize the corpus into structured document data that meets the public opinion monitoring user portrait big model training requirements; S204, the data participants randomly extract a certain amount of training data for annotation as model training samples; S205, the data participants use the Transform architecture to train local sample data within the unit; S206, the data participants obtain the local big model parameters and calculate the Fisher information; S207, the data participants upload the big model parameters and Fisher information to the aggregation server; S208, the aggregation server performs federal resolution of global Fisher information; S209, the aggregation server integrates the local big model parameters uploaded by each data participant into the global big model parameters according to the Fisher information; S210, generates the public opinion monitoring user portrait big model.
6. According to claim 1, the method for automatically monitoring negative public opinion and preventing fraud based on a large model is characterized in that: In S3, public opinion monitoring and anti-fraud include the following steps: S301, real-time collection of public opinion corpus of complaints and comments from Internet or internal company users; S302, cleaning and organizing the corpus into structured document data that meets the training requirements of the public opinion monitoring user portrait big model; S303, inputting the public opinion monitoring user portrait big model for processing; S304, judging whether the current comment constitutes a public opinion event, if not, returning to step S301 to continue the loop; if yes, jumping to the next step to continue execution; S305, judging whether the public opinion event contains fraudulent nature; S306, if the public opinion information is just a hot event and will not cause property and economic losses to users, then the public relations department will pay attention to it simultaneously; S307, if the public opinion information will cause property and economic losses to users, then fix the evidence and synchronize with the legal department.