Negative public opinion information extraction method and device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence and the field of financial science and technology, and discloses a negative public opinion information extraction method, which comprises the following steps of: forming a preliminary data set by acquiring public opinion information data related to a target subject; defining an information extraction structure, labeling the public opinion data set, and generating a training data set; performing multi-round iterative training on the basic pre-training model based on the training data set to generate a negative public opinion information extraction model; and constructing a processing framework based on the model, inputting the to-be-processed public opinion data into the processing framework, automatically identifying negative public opinion information, and generating structured entity, relationship and event data. According to the method, by obtaining the public opinion information data related to the target subject, defining and labeling the information extraction structure and combining the pre-training model of multi-round iterative training, key entities, relationships and events in the public opinion data can be accurately recognized, and automatic extraction of negative public opinion information is achieved. The accuracy and comprehensiveness of negative public opinion information extraction can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology and financial technology, and in particular to a method, device, equipment and storage medium for extracting negative public opinion information. Background Art

[0002] In the financial field, negative public opinion information about enterprises plays an important role in predicting the potential risks of enterprises. Especially in supply chain finance and risk monitoring, negative public opinion monitoring has become a key risk assessment method. By analyzing the public opinion information of enterprises on the Internet, financial institutions can timely discover risk events such as overdue debts and lawsuits that have a significant impact on the operation of enterprises, thereby more comprehensively controlling the credit risks of related enterprises.

[0003] In the new supply chain finance model, banks and financial institutions support the financing needs of upstream and downstream supply chains by approving the credit lines of core enterprises. However, the operating conditions of core enterprises themselves are crucial to the entire supply chain finance system. Once a risk event occurs in a core enterprise, the entire supply chain may be affected. Traditional risk review mainly relies on static data such as financial data, industrial and commercial information, and judicial records. However, these information are updated with a lag and it is difficult to reflect the dynamic risk status of the enterprise in a timely manner. Therefore, negative information monitoring based on Internet public opinion data can effectively make up for the limitations of static data and help financial institutions make timely predictions of the dynamic risks of enterprises.

[0004] Existing corporate public opinion monitoring technologies mostly use regular matching methods to extract news subjects, but this method has obvious limitations. The results of its extraction often include company names that are not related to the article, which interferes with financial institutions when judging corporate risks and makes it impossible to accurately identify the subjects that are truly related to the event. In addition, regular matching methods have difficulty processing subject information that is not in the matching rules, resulting in incomplete subject identification and missing important risk signals.

[0005] The current risk assessment of public opinion monitoring mainly relies on the title or brief description of the event, lacking in-depth analysis of the specific content of the event. In some negative public opinion events, it is difficult to accurately judge the severity of the event and its potential impact on the enterprise based on the title alone. Event analysis should cover a detailed analysis of the event subject, cause, development process and results. These analyses can help financial institutions better judge the impact of the event on corporate risks, but existing technologies have not yet solved this problem. Summary of the invention

[0006] The main purpose of the present invention is to provide a method, device, equipment and storage medium for extracting negative public opinion information, aiming to solve the technical problem that the existing technology is insufficient in the recognition depth of negative public opinion information, resulting in insufficient timeliness and comprehensiveness in enterprise risk monitoring.

[0007] To achieve the above object, the present invention provides a method for extracting negative public opinion information, comprising:

[0008] Acquire public opinion information data related to the target subject, and form a public opinion information data set based on the public opinion information data;

[0009] An information extraction structure is defined according to information extraction requirements, and the public opinion information data set is annotated based on the information extraction structure to generate a training data set, wherein the annotation includes annotation of entities, relationships and events in the public opinion information data;

[0010] Determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information;

[0011] Based on the information extraction model, a processing framework for extracting negative public opinion information is constructed, wherein the processing framework includes an information extraction module;

[0012] Acquire the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework;

[0013] The information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0014] Furthermore, in order to achieve the above-mentioned purpose, the present invention provides a negative public opinion information extraction device, comprising:

[0015] A data acquisition module, used to acquire public opinion information data related to a target subject, and form a public opinion information data set based on the public opinion information data;

[0016] A data annotation module, used to define an information extraction structure according to information extraction requirements, annotate the public opinion information data set based on the information extraction structure, and generate a training data set, wherein the annotation includes annotation of entities, relationships, and events in the public opinion information data;

[0017] A model training module is used to determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information;

[0018] A framework construction module, used to construct a processing framework for extracting negative public opinion information based on the information extraction model, wherein the processing framework includes an information extraction module;

[0019] A data input module, used to obtain the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework;

[0020] An information extraction module is used for the information extraction module of the processing framework to identify negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0021] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and a negative public opinion information extraction program stored in the memory and executable on the processor, and the negative public opinion information extraction program, when executed by the processor, implements the steps of the negative public opinion information extraction method as described above.

[0022] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a negative public opinion information extraction program is stored, and when the negative public opinion information extraction program is executed by a processor, the steps of the negative public opinion information extraction method as described above are implemented.

[0023] Beneficial effects: The present invention relates to the fields of artificial intelligence technology and financial technology, and discloses a method for extracting negative public opinion information. The method forms a preliminary data set by acquiring public opinion information data related to a target subject; defines an information extraction structure, annotates the public opinion data set, and generates a training data set; performs multiple rounds of iterative training on a basic pre-trained model based on the training data set to generate a negative public opinion information extraction model; constructs a processing framework based on the model, and inputs the public opinion data to be processed into the processing framework to automatically identify negative public opinion information and generate structured entity, relationship, and event data. The present invention acquires public opinion information data related to a target subject, defines an information extraction structure and annotates it, and combines a pre-trained model with multiple rounds of iterative training to accurately identify key entities, relationships, and events in public opinion data, thereby realizing the automatic extraction of negative public opinion information. The method can effectively improve the accuracy and comprehensiveness of negative public opinion information extraction, reduce information omissions, and enhance the real-time monitoring capabilities of financial institutions on potential risks of enterprises, thereby enhancing the timeliness of risk warnings and the reliability of decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0025] Figure 1 A schematic diagram of an application environment of a method for extracting negative public opinion information in an embodiment of the present invention;

[0026] Figure 2 This is a flow chart of an embodiment of a method for extracting negative public opinion information of the present invention;

[0027] Figure 3 This is a functional module diagram of a preferred embodiment of the negative public opinion information extraction device of the present invention;

[0028] Figure 4 A schematic diagram of the structure of a computer device in one embodiment of the present invention;

[0029] Figure 5 FIG. 4 is another schematic diagram of the structure of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0030] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0031] The negative public opinion information extraction method provided by the embodiment of the present invention can be applied in Figure 1 In the application environment, the user terminal communicates with the server terminal through the network. The server terminal can obtain the public opinion information data related to the target subject through the user terminal to form a preliminary data set; define the information extraction structure, annotate the public opinion data set, and generate a training data set; perform multiple rounds of iterative training on the basic pre-training model based on the training data set to generate a negative public opinion information extraction model; build a processing framework based on the model, and input the public opinion data to be processed into the processing framework, automatically identify negative public opinion information and generate structured entities, relationships and event data. The present invention can accurately identify key entities, relationships and events in public opinion data by acquiring public opinion information data related to the target subject, defining the information extraction structure and annotating it, and combining the pre-training model of multiple rounds of iterative training, so as to realize the automatic extraction of negative public opinion information. It can effectively improve the accuracy and comprehensiveness of negative public opinion information extraction, reduce information omissions, and enhance the real-time monitoring ability of financial institutions on potential risks of enterprises, thereby enhancing the timeliness of risk warning and the reliability of decision support. Among them, the user terminal can be but not limited to various personal computers, laptops, smart phones, tablet computers and portable wearable devices. The server can be implemented with an independent server or a server cluster composed of multiple servers. The present invention is described in detail below through specific embodiments.

[0032] See also Figure 2 , Figure 2 This is a flow chart of an embodiment of a method for extracting negative public opinion information provided by the present invention. It should be noted that although a logical sequence is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0033] like Figure 2 As shown, the negative public opinion information extraction method proposed by the present invention includes the following steps:

[0034] S10, acquiring public opinion information data related to the target subject, and forming a public opinion information data set based on the public opinion information data;

[0035] In this embodiment, public opinion information related to a specific target subject (such as a company, brand or individual) is obtained from a variety of data sources. Public opinion information may include news reports, social media comments, announcements, blog posts, etc. The target subject is usually a specific company, business or individual, and public opinion information is the public opinion, discussion or news reports around these subjects.

[0036] There are generally two ways to obtain public opinion information data: one is to automatically crawl through web crawlers, and the other is to directly obtain it from third-party data sources through API calls. The types of information obtained include but are not limited to news articles, social media posts, announcements and comments, etc. The purpose is to ensure that the public opinion information obtained can fully reflect the negative, positive and neutral content of the target subject.

[0037] Web crawler: Use customized web crawler programs to automatically crawl relevant public opinion information on the Internet based on the target subject’s keywords. The crawler program can regularly crawl data from different websites to ensure real-time updates of public opinion information.

[0038] API call: extract public opinion data related to the target subject from third-party data sources by accessing open or authorized APIs. For example, use the API of a social media platform to obtain user comments or discussions related to the target subject.

[0039] Conduct a preliminary screening of the acquired data to ensure that only content related to the target subject is retained, and use keyword matching or text analysis methods to filter out irrelevant data.

[0040] After obtaining public opinion information, the next step is to convert these unstructured data into structured public opinion information data sets. The core of this step is to organize, clean, and format public opinion data from different sources to make them into data sets that can be further processed and analyzed. This process ensures the standardization and consistency of the data, providing a good foundation for subsequent information extraction and analysis.

[0041] The processing steps may include:

[0042] Data cleaning: Remove noise data and invalid information, such as HTML tags, advertisements, useless symbols, etc., to make the data easier to process.

[0043] Formatting: Converting data into structured formats, such as JSON or CSV, which facilitates subsequent processing and analysis.

[0044] Deduplication: If there is duplicate content in the data obtained from multiple channels, deduplication is performed to ensure that each piece of information is only retained once in the data set to avoid redundancy.

[0045] The processed public opinion information dataset is stored in the database, and an index is created for each piece of data in the dataset to facilitate subsequent query and analysis operations.

[0046] Example: The bank can regularly obtain public opinion data related to the core enterprise, including news articles, social media discussions, and corporate information in industry reports. Using web crawlers and API calls, the bank's data processing system can capture thousands of relevant data from news websites, industry forums, and social platforms every day. After cleaning and denoising, these data form a clear and unified public opinion information data set, avoiding the interference of redundant data and irrelevant information.

[0047] Banks can then analyze these information data sets to identify negative public opinion information and mine potential risk signals, such as debt defaults, financial distress, or negative social news. Based on these data sets, banks can issue timely warnings, assess the operational risks of core enterprises, and adjust their credit lines to avoid the occurrence of bad debts.

[0048] By obtaining public opinion information related to the target subject from multi-source data and organizing it into a standardized public opinion information data set, subsequent data processing and analysis will be more efficient and accurate.

[0049] S20, defining an information extraction structure according to information extraction requirements, annotating the public opinion information data set based on the information extraction structure, and generating a training data set, wherein the annotation includes annotations of entities, relationships, and events in the public opinion information data;

[0050] In this embodiment, according to the specific needs of public opinion information analysis, an information extraction structure is defined to annotate key information (such as entities, relationships, and events) in public opinion data. The information extraction structure refers to the standards and rules for identifying specific types of information from public opinion information. These standards include the types of entities (such as company names, personal names), relationships (such as contractual relationships, litigation relationships), and events (such as debt defaults, litigation, etc.).

[0051] Define the entity categories that need to be identified according to different public opinion needs. For example, in corporate public opinion, entities can be company names, people, places, etc.

[0052] According to the relationship between different entities, define the relationship type in the information extraction structure, for example, the relationship between an enterprise and a contract dispute, the employment relationship between an individual and an enterprise, etc.

[0053] Define the types of events that need to be identified based on analysis requirements. For example, events in public opinion can be bankruptcy, litigation, debt default, etc.

[0054] Set standards for each information extraction structure to ensure that when the model processes the data, it can clearly extract the required entity, relationship, and event information. Natural language processing (NLP) technology can be used to automatically extract this information.

[0055] According to the information extraction structure defined above, each data item in the public opinion information dataset is labeled. Labeling means marking each information item in the dataset as a certain type of entity, relationship or event according to a predefined structure. Labeling can be done through automated tools or human assistance to generate training datasets for model learning.

[0056] You can use natural language processing-based annotation tools to automatically annotate entities, relationships, and events in a dataset. For example, you can use named entity recognition (NER) technology to annotate company names, key figures, etc. in a text, and use dependency parsing to identify relationships between entities.

[0057] After automatic annotation, manual annotation correction can be performed to ensure the accuracy and completeness of the automatic annotation. Manual correction is particularly important for complex relationship and event annotations.

[0058] During the annotation process, multi-level annotation may be performed, for example, while annotating the entities, the relationships involved in the events are also annotated, and the time and place of the events are further annotated.

[0059] The annotated public opinion information dataset will be used to generate a training dataset, which is labeled data for the model to learn and train. When generating a training dataset, ensure that the dataset covers comprehensive annotations of entities, relationships, and events so that the model can learn how to extract information from these labels.

[0060] The labeled data is stored by entity, relationship, and event categories to ensure that all key information is correctly labeled. The data is organized into a format suitable for machine learning model training, such as JSON, CSV, or other standard data formats to ensure that the model can directly use this data for training. The labeled data is stored in the training dataset library and prepared for subsequent multiple rounds of iterative training to ensure data integrity and consistency.

[0061] Example: In the financial field, suppose a bank needs to monitor the public opinion information of its core enterprises to prevent risk events from happening to the enterprises. The bank can obtain news and social media data related to the enterprise from the Internet, and automatically annotate information such as the enterprise name and debt disputes according to the defined information extraction structure. Finally, the generated training data set is used to train the public opinion analysis model to help the bank identify potential risk events, such as debt defaults or legal proceedings.

[0062] By defining a clear information extraction structure, we can accurately identify entities, relationships and events in public opinion information, and generate high-quality training data sets by combining automatic annotation with manual correction. Compared with existing technologies, the annotation process is more efficient and accurate, and the generated training data sets cover complete public opinion information, which can provide an effective training data source for the model, thereby improving the accuracy and comprehensiveness of negative public opinion information extraction.

[0063] S30, determining a basic pre-training model, performing multiple rounds of iterative training on the basic pre-training model based on the training data set, updating and optimizing parameters of the basic pre-training model, and generating an information extraction model for extracting negative public opinion information;

[0064] In this embodiment, the basic pre-trained model is a natural language processing model that has been trained with a large-scale general data set, such as BERT, GPT, etc. These models have a basic understanding of language by learning a large amount of general language data, but are not optimized for specific tasks. Therefore, in this step, you first need to select a basic pre-trained model as the basic model for the subsequent public opinion information extraction task.

[0065] According to the needs of public opinion information, select a suitable basic pre-training model. Commonly used models include BERT (for semantic understanding and classification tasks), GPT (for generation and language reasoning tasks), and ERNIE (for knowledge-enhanced entity recognition and information extraction tasks). Load the selected basic pre-training model into the system as the starting point for subsequent training. Perform a basic evaluation of the selected pre-training model to ensure that it is suitable for the extraction of negative public opinion information and can be well combined with the training dataset.

[0066] In addition, you can also choose the UIE model. The UIE (Universal Information Extraction) model is a model based on the ERNIE3.0 knowledge enhancement pre-training model and optimized for information extraction tasks. The main advantage of UIE is that it can perform various information extraction tasks such as entities, relationships, and events through a unified structured extraction language (SEL). Unlike traditional information extraction models, UIE unifies the extraction process of different tasks to model and thus has the ability to migrate and generalize between tasks. The UIE model is particularly suitable for the following scenarios:

[0067] Small sample fine-tuning: The UIE model can handle situations with insufficient data or sparsely labeled samples during training and has efficient learning capabilities.

[0068] Fast cold start: Since the model has good generalization ability for different information extraction tasks, UIE can start quickly with a small amount of data or task changes, and adapt to different extraction targets.

[0069] Multi-task joint modeling: UIE can handle multiple tasks simultaneously, such as entity extraction, relationship collection, and event extraction, and encode them through a unified extraction framework, reducing the complexity of training different task models.

[0070] Knowledge enhancement: Based on ERNIE3.0, UIE can combine a large amount of external knowledge to enhance its understanding of entities and relationships within the domain, especially in professional fields (such as finance, law, etc.), with higher recognition accuracy.

[0071] Once the basic pre-trained model is selected, it needs to be trained in a targeted manner. The training process is based on the labeled training data set and fine-tuning the model through multiple training rounds. The purpose of multiple rounds of iterative training is to allow the model to learn specific task features from the training data, such as identifying entities, relationships, and events in negative public opinion information.

[0072] The labeled training data set is input into the model. The data set contains labeled data such as entities, relationships, and events of public opinion information. The training process uses multiple rounds of iterative training. In each round of iteration, the model adjusts parameters to improve its understanding and extraction capabilities of specific tasks. The training process can use the back propagation algorithm to correct errors and gradually optimize the weights of the model. During the training process, different hyperparameters are set, such as learning rate, batch size, training rounds, etc., to ensure the effectiveness and convergence of the training.

[0073] In each round of iterative training, the model gradually optimizes its internal parameters by learning the training data set. The purpose of parameter optimization is to make the model better adapt to the task of public opinion information extraction, including identifying key entities, relationships, and events in negative public opinion. The updated and optimized model will have a higher accuracy in identifying negative public opinion.

[0074] Through multiple rounds of iterative training and parameter optimization, the basic pre-training model is gradually transformed into a specific information extraction model for the task of extracting negative public opinion information. The model can automatically identify entities, relationships and events in public opinion information, especially information related to negative public opinion, based on the features learned during the training process.

[0075] After training, the model parameters are optimized, and the generated model has the ability to extract negative public opinion information. At this point, the model has the ability to analyze and extract the input public opinion data. The trained model is saved to the system for use in the actual negative public opinion information extraction task. The model can be deployed to the information extraction processing framework to process the input public opinion data in real time and identify negative public opinion information.

[0076] Example: In financial risk monitoring, banks need to continuously monitor public opinion information of credit customers, especially negative information involving debt defaults, lawsuits, etc. By adopting the UIE model, banks can quickly deploy a system for extracting negative public opinion information and identify negative entities, events, and relationships related to customers from a large amount of news, announcements, and social media data. The advantage of the UIE model lies in its fast cold start capability. It can start quickly with a small amount of labeled data, and use the knowledge enhancement function of ERNIE3.0 to accurately identify proprietary entities in the financial field, such as company names, debt default events, etc.

[0077] Through multiple rounds of iterative training and optimization of the basic pre-trained model, a model specifically used for extracting negative public opinion information can be generated. This model can accurately identify entities, relationships, and events in public opinion information, and has a high extraction accuracy. Compared with existing public opinion information extraction methods, the generated information extraction model is more suitable for negative public opinion analysis, can identify potential risks in a timely manner, help financial institutions, enterprises, etc. to respond and make decisions in a timely manner, and improve the accuracy and response speed of risk warnings.

[0078] S40, constructing a processing framework for extracting negative public opinion information based on the information extraction model, wherein the processing framework includes an information extraction module;

[0079] In this embodiment, based on the previously generated and fine-tuned public opinion information extraction model, a processing framework capable of processing and managing negative public opinion information is built. The processing framework is used to process the input public opinion data through specific steps and finally output negative public opinion information. The construction of the framework mainly revolves around the integration of the information extraction model to ensure that the model can stably and efficiently perform information extraction tasks within the framework.

[0080] In view of the characteristics of negative public opinion information, an overall framework structure with data reception, processing and output is designed to ensure that the extraction model can be smoothly integrated and run. The processing framework adopts a modular design, and the information extraction model plays a role in the framework as a core component. Other parts of the framework such as data input, data preprocessing and result output will support and serve the operation of the information extraction model. The information extraction model that has been fine-tuned in the early stage is embedded in the framework to ensure that the model can smoothly receive the input data within the framework and perform efficient extraction of negative public opinion information.

[0081] The information extraction module is the core functional module of the processing framework. It is used to analyze the input data through the model and extract negative entities, events and relations in the public opinion information. This module is responsible for calling the previously trained public opinion information extraction model and performing information recognition and extraction tasks according to the structure of the model. The fine-tuned negative public opinion information extraction model is configured as the main execution unit in the framework. This module receives the input data in the framework and extracts information through the extraction model. The information extraction module in the framework calls the loaded negative public opinion information extraction model to identify key information in the input data, such as negative entities, events and relations. During the operation of the information extraction module, the model is used to analyze the input public opinion data, extract negative information and pass it to the subsequent processing module.

[0082] By building a negative public opinion information processing framework that includes an information extraction module, we can efficiently integrate a fine-tuned negative public opinion information extraction model to achieve automated processing of negative public opinion data. The framework has high scalability and integration capabilities, and can flexibly call information extraction models according to different public opinion data sources and task requirements, thereby improving the recognition accuracy and analysis efficiency of negative public opinion information.

[0083] S50, obtaining the public opinion data to be processed related to the target subject, and inputting the public opinion data to be processed into the processing framework;

[0084] In this embodiment, public opinion data related to the target subject is collected from different information sources. The target subject is usually a specific enterprise, organization or individual, and its public opinion data may include news reports, announcements, discussions and comments on social media, and other public information. The public opinion data to be processed needs to be obtained through multiple channels, usually including the following sources:

[0085] News sources: Includes reports from traditional media and online news platforms.

[0086] Announcement source: announcements issued by companies or institutions, especially those involving finance, operations, law, etc.

[0087] Social media: discussions and comments from social platforms (such as Weibo, Twitter, Facebook, etc.), especially consumer feedback and market opinion.

[0088] This data may be raw and unprocessed and needs to be further cleaned, formatted and processed before it can be used for information extraction.

[0089] Through web crawlers or API interfaces, data related to the target subject is regularly collected from the above multi-source information. According to predefined keywords or identification rules, public opinion information directly related to the target subject (such as company name, brand, business, etc.) is screened out to form a public opinion data set to be processed.

[0090] After collecting and filtering out the public opinion data to be processed related to the target subject, these data will be input into the processing framework. The processing framework is the core part of the entire system, which is responsible for receiving the input data and performing subsequent processing and analysis. The input data format may include plain text, HTML, PDF, pictures, etc. The processing framework needs to support the input of multiple data formats.

[0091] Convert public opinion data in different formats (such as text, images, PDF, etc.) into a data format suitable for the processing framework, such as plain text format, so that the subsequent steps can process the data. Input the processed public opinion data into the processing framework through an interface or API. The structured design of the processing framework enables it to receive, identify and classify different types of public opinion data, preparing for subsequent preprocessing, information extraction and other steps.

[0092] Through the integration of multi-source information collection and processing frameworks, public opinion data related to the target subject can be automatically acquired and processed. Not only can public opinion data be extracted from multiple information sources, but the data can also be formatted and input into the framework for subsequent processing, realizing the automatic identification and output of negative public opinion information. Through this automated process, the efficiency and accuracy of public opinion information processing are significantly improved.

[0093] S60, the information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0094] In this embodiment, in the processing framework, the information extraction module is the core module, which is responsible for identifying negative public opinion information in the public opinion data to be processed through the defined information extraction structure. This process includes the identification of entities, relationships, and events. Negative public opinion information specifically refers to information that has an adverse impact on the target subject (such as a company, brand, or individual), usually involving some negative events or activities, such as financial crisis, legal disputes, negative reports, etc.

[0095] The information extraction structure defines the types of information that the system needs to recognize and extract, such as entities (company name, event date, location, etc.), relationships (such as the relationship between a company and the litigation object), and events (such as debt default, litigation, etc.). The information extraction structure is a template or rule designed in advance and used to guide the information extraction process.

[0096] Entity recognition mainly extracts core entities from text, such as company names, brands, time of event occurrence, etc. Through the pre-trained model, the information extraction module can accurately identify these entities.

[0097] Relationship identification is responsible for identifying the associations between entities. For example, whether a company has a legal dispute with another company, or how a certain event affects the company. Through the relationship templates in the information extraction structure, the system can identify and extract this relationship information.

[0098] Event identification is to extract important events related to the target subject from public opinion data, such as the company's legal disputes, financial difficulties, market crises, etc. These events are key information for negative public opinion analysis and directly affect the company's reputation and market performance.

[0099] First, the system loads a trained and optimized negative public opinion information extraction model, which is fine-tuned based on specific training data to identify specific types of entities, relationships, and events. The public opinion data to be processed is input into the information extraction module, and the data will undergo preliminary preprocessing, such as denoising, formatting, etc. According to the information extraction structure, the module analyzes the input data step by step. First, identify entities (such as company name, time, location, etc.), then identify the relationships between these entities (such as cooperative relationships, legal relationships, etc.), and finally identify events related to these entities (such as lawsuits, negative news, etc.). The module generates structured negative public opinion information, including the identified entities, relationships, and events, and uses this information as input for subsequent analysis.

[0100] The information extraction module can efficiently identify negative public opinion information in the public opinion data to be processed, and accurately extract key information such as entities, relationships and events. The information extraction module can automatically identify negative events of the target subject, help enterprises and financial institutions to understand market changes and potential risks in a timely manner, and provide data support for subsequent decision-making. Through this systematic and automated public opinion processing, the accuracy and efficiency of negative public opinion analysis can be greatly improved.

[0101] The present invention relates to the fields of artificial intelligence technology and financial technology, and discloses a method for extracting negative public opinion information. The method forms a preliminary data set by acquiring public opinion information data related to a target subject; defines an information extraction structure, annotates the public opinion data set, and generates a training data set; performs multiple rounds of iterative training on a basic pre-training model based on the training data set to generate a negative public opinion information extraction model; constructs a processing framework based on the model, and inputs the public opinion data to be processed into the processing framework, automatically identifies negative public opinion information, and generates structured entity, relationship, and event data. The present invention acquires public opinion information data related to a target subject, defines an information extraction structure and annotates it, and combines a pre-training model with multiple rounds of iterative training to accurately identify key entities, relationships, and events in public opinion data, thereby realizing the automatic extraction of negative public opinion information. The method can effectively improve the accuracy and comprehensiveness of negative public opinion information extraction, reduce information omissions, and enhance the real-time monitoring capabilities of financial institutions on potential risks of enterprises, thereby enhancing the timeliness of risk warnings and the reliability of decision support.

[0102] In one embodiment, the above S10 includes:

[0103] S101, performing text denoising processing on the public opinion information data;

[0104] S102, performing word segmentation processing on the processed public opinion information data, dividing the text in the public opinion information data into single words or phrases;

[0105] S103, performing sentiment analysis on the text after word segmentation processing, determining the sentiment tendency of the text, and marking the text with a corresponding sentiment tendency label;

[0106] S104, generating the public opinion information dataset based on the text annotated with the sentiment tendency label.

[0107] In this embodiment, public opinion information data related to the target subject (such as a company, brand, or individual) is obtained from multiple information sources (such as news, announcements, and social media). The data set needs to cover all information related to the target subject, whether it is positive, negative, or neutral. The establishment of the public opinion information data set is intended to provide basic data for subsequent processing and analysis.

[0108] First, define and screen out the main information sources related to the target subject, such as news websites, social media, forums, announcements, etc. Automatically collect public opinion data related to the target subject from these information sources through web crawlers, API interfaces and other technical means. Format the collected data and unify it into a standardized text data format suitable for subsequent analysis.

[0109] The purpose of text denoising is to clean up invalid information in the original data, such as extra blank characters, redundant symbols, advertising information, etc., so as to improve the accuracy of subsequent data processing and analysis. Denoising can effectively remove irrelevant parts of public opinion data and ensure data cleanliness.

[0110] Delete useless symbols, HTML tags, blank characters, etc. in the text. Remove irrelevant advertisements, duplicate content or invalid information to ensure the validity of the data.

[0111] Word segmentation is a basic step in natural language processing (NLP), which aims to segment continuous text into individual words or phrases for subsequent analysis and processing. Since public opinion data is often unprocessed natural language, word segmentation can help the system identify and extract key vocabulary and semantic information.

[0112] Use rule-based or statistical word segmentation algorithms (such as dictionary-based word segmentation and machine learning word segmentation) to segment the text. While segmenting, perform part-of-speech tagging on the words to provide a grammatical basis for subsequent sentiment analysis or entity recognition.

[0113] The purpose of sentiment analysis is to determine the sentiment tendency of a text, including positive, negative or neutral, based on the semantics expressed in the public opinion data. Sentiment analysis can help the system understand the emotional state in public opinion information and provide important clues for subsequent negative public opinion identification. The sentiment tendency label of the text will be used for further data classification and analysis.

[0114] Use a pre-trained sentiment analysis model (such as a machine learning or deep learning-based model) to classify the sentiment of the text. Based on the sentiment analysis results, automatically label the text with sentiment (such as "positive", "negative" or "neutral").

[0115] After sentiment analysis of the text and labeling of sentiment tendency, a public opinion information dataset is generated. This step integrates all sentiment analysis and labeled text data to form a complete dataset for subsequent information extraction and analysis. The dataset is highly structured, which facilitates further analysis and processing by the system.

[0116] The texts marked with sentiment labels are organized into a unified data format and integrated into a public opinion information dataset. The generated dataset is stored in a specific file format (such as CSV, JSON) for subsequent public opinion analysis or information extraction steps.

[0117] This embodiment can efficiently generate a public opinion information data set marked with sentiment by performing text denoising, word segmentation and sentiment analysis on public opinion information data. This not only improves the processing accuracy of public opinion data, but also helps users quickly identify negative public opinion information, providing a reliable data basis for risk warning and public opinion monitoring. Through this automated public opinion data processing, the efficiency and accuracy of public opinion monitoring can be significantly improved.

[0118] In one embodiment, in the above S20, an information extraction structure is defined according to information extraction requirements, and the public opinion information dataset is annotated based on the information extraction structure to generate a training dataset, including:

[0119] S201, defining an information extraction structure according to information extraction requirements, wherein the information extraction structure includes identification criteria for entities, relationships, and events in public opinion information data;

[0120] S202, labeling entities in the public opinion information dataset based on entity identification criteria;

[0121] S203, annotating the relationships in the public opinion information dataset based on the relationship identification standard;

[0122] S204, marking the events in the public opinion information dataset based on the event identification standard;

[0123] S205, generating the training data set based on the annotated public opinion information data set.

[0124] In this embodiment, in order to set clear standards and rules so that target entities, relationships and events can be accurately identified in the subsequent public opinion information data processing, the definition of the information extraction structure is the core step, which determines how the system identifies and processes different categories of information.

[0125] Entities can be key nouns related to the target subject, such as company names, people, places, dates, etc. Relationships are interactions and associations between entities, such as the cooperative relationship between a company and its suppliers, or the legal relationship between a company and the party being sued. Events refer to specific behaviors or activities described in public opinion information, such as a company's financial crisis, lawsuits, negative news, etc.

[0126] Based on domain knowledge and analysis requirements, define which entity types need to be extracted and annotated. According to the possible interactions and relationships between different entities, set corresponding rules to identify the associations between entities. Clarify the event types and standards that need to be extracted in public opinion information in order to identify and extract important events related to the target subject.

[0127] After defining the entity recognition criteria in the information extraction structure, the system needs to label the entities in the public opinion information dataset according to these criteria. Entity labeling refers to identifying specific nouns or keywords in the text and marking them as preset entity categories. For example, company names, key figures, event times, etc. can all be labeled as entities.

[0128] Using natural language processing techniques (such as named entity recognition, NER), entities are automatically extracted from text according to preset criteria. Based on the recognition results, the system labels each entity and classifies it into predefined entity categories, such as companies, people, dates, etc.

[0129] Relationship identification means that after the system identifies the entities, it further analyzes the relationships between entities. Relationships can be business cooperation, legal proceedings, competitive relationships, etc. The purpose of this step is to extract the interaction information between entities and annotate the relationships.

[0130] Using relationship extraction technology, we analyze the interaction patterns between different entities in public opinion texts and identify the specific relationships between them. We classify the identified relationships and mark them into predefined categories such as cooperative relationships, legal relationships, and business relationships.

[0131] Event labeling is one of the key steps in information extraction, which is responsible for identifying important events in public opinion information. Events are usually key behaviors or status changes of the target subject in news, announcements or social media, such as financial problems, legal proceedings, etc. The system needs to label relevant information as specific event categories based on event identification criteria.

[0132] Use the event extraction model to extract event information by identifying keywords and patterns in the text. Based on the identified event information, it is labeled as a specific event category, such as "financial crisis", "litigation event", "negative news", etc.

[0133] After completing the annotation of entities, relationships, and events, the system integrates these annotated data to generate a structured training data set. This data set will serve as the training basis for the information extraction model to help the model better recognize future input data.

[0134] Integrate the annotated entity, relationship, and event information into a structured dataset. Generate the dataset required for model training based on the annotated dataset, ensuring that the quality and format of the dataset meet the training requirements.

[0135] By defining the information extraction structure, this embodiment can efficiently identify key entities, relationships and events in public opinion information and generate high-quality training data sets. It not only improves the training efficiency of the model, but also significantly improves the model's ability to identify negative public opinion information. Through this automated labeling and data set generation, the system can more accurately extract negative public opinion information related to the enterprise, helping enterprises and financial institutions to achieve real-time public opinion monitoring and risk warning.

[0136] In one embodiment, the above S30 includes:

[0137] S301, determining a pre-trained model that has been trained on a general text dataset;

[0138] S302, setting training parameters, and performing multiple rounds of iterative training on the pre-trained model based on the training parameters and the training data set;

[0139] S303, adjusting and optimizing the parameters of the pre-trained model by a back propagation algorithm in each round of training;

[0140] S304, when multiple rounds of iterative training are completed, the information extraction model for extracting negative public opinion information is obtained.

[0141] In this embodiment, a basic pre-trained model trained with a general data set is selected as the initial model for extracting negative public opinion information. The pre-trained model is usually pre-trained based on large-scale general text data (such as Wikipedia, news text, etc.), which can learn a wide range of language patterns and knowledge and has certain language understanding and information extraction capabilities. Based on this, targeted data fine-tuning and optimization will be carried out later.

[0142] Commonly used pre-trained models include BERT, ERNIE, etc. These models have been trained extensively on common data sets and have good language understanding capabilities. The selected pre-trained model is loaded into the system and fine-tuned and optimized as a base model.

[0143] In order to adapt the basic pre-trained model to the needs of extracting negative public opinion information in a specific field, the model needs to be trained for multiple rounds of iterations. In this process, the training parameters are first set, including learning rate, batch size, number of training rounds, etc. The training dataset is annotated domain-specific data that can help fine-tune the model in identifying negative public opinion information and further improve its performance in a specific field.

[0144] Training parameters include learning rate (determines the step size of each weight update of the model), batch size (the amount of data input into the model for training each time), and number of training rounds (the number of times the entire dataset is trained). Based on the set training parameters, the model is trained for multiple rounds to ensure that the model continues to learn the patterns and features in the training data.

[0145] Backpropagation is the core algorithm for deep learning model training. It is used to calculate the gradient of the loss function and optimize the model parameters through the gradient descent method. In each round of training, the model makes predictions based on the input training data, compares the predicted results with the actual results, and calculates the error. Through the backpropagation algorithm, the system will gradually adjust the model parameters to reduce the prediction error and improve the accuracy of the model.

[0146] Input training data, the model makes predictions, and obtains prediction results. Compare the prediction results with the true labels and calculate the loss value. The gradient of the loss to the parameters of each layer of the model is calculated through the back propagation algorithm, and the parameters are gradually updated to reduce the loss value and optimize the model performance.

[0147] After multiple rounds of iterative training and optimization, the model parameters are gradually adjusted to the optimal state. At this time, the model has the ability to identify negative public opinion information and becomes a dedicated model for extracting negative public opinion information. This model can be used for real-time public opinion data processing and analysis to help enterprises or financial institutions identify potential negative information.

[0148] After multiple rounds of training, the system determines whether the training has converged based on preset conditions (such as the performance indicators of the validation set). The trained negative public opinion information extraction model is saved for subsequent practical applications and reasoning.

[0149] This embodiment effectively improves the model's ability to identify negative public opinion information by selecting a pre-trained model trained with a general text dataset and combining it with a negative public opinion dataset in a specific field for multiple rounds of iterative training. The generated negative public opinion information extraction model can accurately identify entities, relationships, and events in public opinion data through parameter optimization using the back-propagation algorithm, providing enterprises or financial institutions with efficient and accurate public opinion analysis tools to help them respond to potential risks in a timely manner.

[0150] In one embodiment, the above S40 includes:

[0151] S401, configuring a data input module of the processing framework, wherein the data input module is used to receive public opinion information data to be processed;

[0152] S402, configuring a preprocessing module of the processing framework, wherein the preprocessing module is used to perform a preprocessing operation on the input public opinion information data;

[0153] S403, configuring the information extraction model as an information extraction module of a processing framework;

[0154] S404, configuring a feedback learning module of the processing framework, wherein the feedback learning module is used to analyze the difference between the current negative public opinion information and the historical negative public opinion data according to the extraction result of the information extraction module, determine the change pattern including event feature change, sentiment tendency change, entity relationship change and event frequency change, and store the change pattern as feedback data;

[0155] S405, configuring a dynamic update module of the processing framework, wherein the dynamic update module is used to analyze the feedback data stored in the feedback learning module and adjust the parameters of the information extraction model based on the analysis results.

[0156] In this embodiment, the role of the data input module is to receive external public opinion information data, which can come from different sources, such as news, announcements, social media, etc. The main function of this module is to ensure that the data can be received by the system and passed to the subsequent pre-processing module for processing.

[0157] Determine the source of data input, including real-time data streams (news, social media) or historical data storage (announcements, news archives). Set up appropriate data transmission protocols and interfaces (such as APIs, database connections, message queues, etc.) to ensure that the public opinion information data to be processed can be efficiently input into the system.

[0158] The preprocessing module is one of the core components of the processing framework, responsible for a series of preliminary cleaning and standardization of the input public opinion information data. Preprocessing operations include text denoising, word segmentation, deduplication, formatting, etc. to ensure that the data can meet the subsequent processing requirements.

[0159] Remove noise data such as irrelevant symbols and blank characters from the text. Split continuous natural language text into recognizable words or phrases to facilitate subsequent analysis and processing. Check for duplicate information in the input data set and delete duplicate entries to ensure the uniqueness of each piece of data. Convert the processed data into a unified format (such as JSON, CSV) to ensure the consistency of the data structure for the operation of subsequent modules.

[0160] The information extraction model is the core module in the processing framework. Its task is to extract key entities, relationships, events and other information from the input public opinion information data according to the pre-trained model. This module will use the previously trained negative public opinion information extraction model to automatically identify and extract valuable public opinion information from the data.

[0161] Load the trained information extraction model into the system and use it as the core of the information extraction module. According to the input data format and processing requirements of the system, configure the input and output interfaces of the information extraction model to ensure that the input data can be processed by the model and generate structured output results (such as entities, relationships, event information, etc.). The information extraction module will use the model to identify important entities, relationships, and events based on the input public opinion data, and generate corresponding structured information for subsequent analysis.

[0162] The core function of the feedback learning module is to capture changes in negative public opinion information in real time, and dynamically identify important change features by comparing the current extraction results with historical data. This module specifically includes the following functions:

[0163] Data comparison: compare the currently extracted negative public opinion information with the stored historical negative public opinion data to identify the differences. Historical data is usually stored in a structured database for subsequent analysis.

[0164] Identify change characteristics: The feedback learning module identifies four main change characteristics: Event characteristic changes: Detect changes in event content (such as event subject, action type, etc.). For example, if a company's negative event changes from litigation to bankruptcy, this change is an event characteristic change. Sentiment tendency changes: Compare the sentiment labels of current and historical public opinion information to identify changes in sentiment, such as from "negative" to "extremely negative." Entity relationship changes: The relationship between entities has changed, such as the litigation relationship between companies evolving into a merger and acquisition relationship. Event frequency changes: Detect whether the frequency of event occurrence has increased or decreased.

[0165] Storing change patterns: The identified change features are stored in a structured format, and this data can be used for further analysis and model optimization.

[0166] The feedback learning module can be implemented through machine learning algorithms, such as time series analysis, cluster analysis, and difference detection algorithms. The specific implementation details of each change feature can be extracted through pre-defined rules or dynamically learned feature patterns.

[0167] The dynamic update module adaptively optimizes the information extraction model based on the change patterns identified by the feedback learning module. The core function of this module is to dynamically adjust the recognition criteria and parameters of the model according to the change pattern data to ensure that the model can accurately adapt to changes in public opinion information.

[0168] The dynamic update module obtains the change pattern data from the feedback learning module and performs a comprehensive analysis. This includes weight assignment and correlation analysis of the change data. After analyzing the change pattern, the dynamic update module first optimizes the recognition criteria of the information extraction module. The process of adjusting the recognition criteria includes the following aspects:

[0169] Sensitivity of event extraction: Adjust the sensitivity of identifying specific events based on feedback data. For example, for certain event types (such as legal disputes or financial changes), the sensitivity of their extraction can be enhanced by lowering the threshold.

[0170] Sentiment analysis weight: In response to changes in sentiment tendencies, the model adjusts the weight ratio of sentiment analysis to make it pay more attention to events with obvious changes in sentiment, especially the transition from "neutral" to "negative" or "negative" to "extremely negative".

[0171] Update of entity association rules: During the entity matching process, the association rules between entities are dynamically adjusted, and these rules are optimized in a timely manner based on feedback data to cope with complex and dynamic public opinion backgrounds.

[0172] The dynamic update module also needs to optimize the parameters of the information extraction model to improve the model's adaptability and accuracy. The specific contents of parameter optimization include:

[0173] Adaptive adjustment of learning rate: Adjust the learning rate of the information extraction model according to the change pattern data. For example, when the change pattern is obvious, increase the learning rate to enhance the model's ability to learn new changes.

[0174] Dynamic optimization of feature weights: The dynamic update module adjusts the weights of different features based on the change features obtained by the feedback learning module. For example, a higher weight is given to "event frequency change" to deal with high-frequency risk events.

[0175] Adjustment of model regularization parameters: When negative public opinion changes significantly, the generalization ability of the model can be improved by adjusting the regularization parameters to avoid overfitting of the model due to large changes.

[0176] The dynamic update module can be implemented by combining rule-based update methods and deep learning-based adaptive optimization. The rule update part can rely on business preset rules, while deep learning optimization can use reinforcement learning, online learning and other methods to continuously improve the adaptability of the model.

[0177] Example: In the public opinion analysis and risk assessment of the core enterprise "Y Company", the bank monitors the public opinion changes of "Y Company" through the negative public opinion information extraction system. Recently, "Y Company" was reported by many media to be involved in a major contract dispute. The feedback learning module of the system analyzed the latest public opinion information and found that the "event feature changes" and "entity relationship changes" of the event were more significant after comparing it with historical data.

[0178] Therefore, the dynamic update module automatically adjusts the model parameters of the system:

[0179] The sensitivity of events involving "contract disputes" has been raised to a high priority, and the extraction threshold for other common public opinion events has been lowered to ensure a quick response.

[0180] For "legal proceedings" events, the weight of sentiment analysis has been increased, and the focus on "negative" to "extremely negative" has been strengthened.

[0181] The learning rate and regularization parameters of the information extraction model were adjusted so that the model can adapt to changes and more accurately identify the correlation between major public opinion events.

[0182] Through this adaptive adjustment method, the system can not only capture potential major legal disputes of "Company Y" in real time, but also adaptively optimize the model's identification criteria and parameters according to the latest changes, providing a more comprehensive and accurate enterprise risk early warning capability.

[0183] This embodiment realizes the automatic processing and analysis of negative public opinion information by configuring the data input module, preprocessing module, information extraction module, feedback learning module and dynamic update module of the processing framework. The data input module ensures the real-time reception of data, the preprocessing module improves the data quality, and the information extraction module efficiently and accurately identifies the key entities, relationships and events in the public opinion data. It can also adaptively optimize the recognition criteria and parameters of the model according to the latest changes, providing a more comprehensive and accurate enterprise risk early warning capability.

[0184] In one embodiment, the above S60 includes:

[0185] S601, performing a preprocessing operation on the public opinion data to be processed by a preprocessing module of the processing framework;

[0186] S602, inputting the public opinion data to be processed after the preprocessing operation into the information extraction module;

[0187] S603, the information extraction module analyzes the public opinion data to be processed according to the information extraction structure, and identifies negative public opinion information containing entity information, relationship information and event information in the public opinion data to be processed;

[0188] S604: Generate target structured information based on the negative public opinion information including entity information, relationship information and event information.

[0189] In this embodiment, the input public opinion data to be processed is preprocessed to ensure that the data quality meets the requirements of information extraction. The task of the preprocessing module is to clean, segment, remove duplicates and format the original data to ensure the consistency and accuracy of the data.

[0190] Remove invalid characters and symbols from public opinion data, such as spaces, HTML tags, and irrelevant special symbols. Split continuous text into independent words or phrases for subsequent analysis. Detect duplicate text entries in the data set and delete duplicate parts to ensure data uniqueness. Process the text into a unified format for subsequent module processing (such as JSON or a specific structured table format).

[0191] After the data is processed by the preprocessing module, it is passed to the information extraction module. The purpose of this step is to pass the cleaned and formatted data to the model so that the model can analyze the data and extract information.

[0192] Pass the pre-processed public opinion data to the information extraction module according to the system interface requirements (such as API interface or data flow). Ensure that the incoming data format is consistent with the model input requirements and can be correctly parsed and processed.

[0193] The core function of the information extraction module is to analyze the input data and extract key information based on the defined extraction structure (such as the identification criteria of entities, relationships, and events). This step will identify the entity, relationship, and event information related to negative public opinion from the public opinion data.

[0194] According to the pre-trained model and extraction structure, extract entity information such as company name, person, time, location, etc. from the data. Extract the relationship between entities, such as cooperative relationship, competitive relationship, legal proceedings, etc. Identify specific events, such as lawsuits, negative news, financial crises, etc., and mark them as negative public opinion events.

[0195] After negative public opinion information is successfully identified and extracted, the system will generate target structured information based on the identified entities, relationships and event information. These structured data can be used for subsequent analysis, display or further processing to help enterprises or financial institutions conduct risk assessment and decision support.

[0196] Organize the identified entities, relationships, and event information into unified structured data (such as tables, JSON format, etc.) and output the results. Store the structured public opinion information in the database for subsequent analysis and monitoring.

[0197] This embodiment can automatically identify negative public opinion information in the data by preprocessing, analyzing and extracting public opinion data, and generate structured information including entities, relationships and events. This processing flow significantly improves the efficiency of extracting public opinion information, helps enterprises and financial institutions to timely discover potential risks, and provides comprehensive decision-making support.

[0198] In one embodiment, after the above S60, the method further includes:

[0199] S701, determining a core business entity related to the target entity;

[0200] S702, performing matching analysis on entities in the public opinion event through an entity matching module to determine a target public opinion event that matches the core business entity;

[0201] S703, analyzing the severity of the target public opinion event based on the entities and relationships in the target public opinion event;

[0202] S704: Generate a corresponding public opinion resolution strategy based on the severity of the target public opinion event.

[0203] In this embodiment, the core business entities directly related to the target entity (such as an enterprise) are clearly defined. These entities may include the target enterprise's main products, services, partners or key figures, usually key roles or elements in the target entity's business operations. Define the scope of the core business entities based on the business characteristics and industry standards of the enterprise. By analyzing the public information of the target enterprise (such as financial reports, announcements, etc.), determine the specific information of the core entities, such as the company name, key partners, main products and services, etc.

[0204] The entity matching module is responsible for matching and analyzing entities in public opinion events to determine whether the entities in the event match the core business entities of the target entity. If the match is successful, the public opinion event is considered an important target public opinion event.

[0205] Based on technologies such as text matching and semantic analysis, matching analysis is performed on entities in public opinion data (such as company names, products, etc.) to determine whether they match core business entities. A matching threshold, such as similarity or relevance score, is set to ensure that only entities that are highly relevant to the core business are considered as matching target public opinion events. Based on the matching results, public opinion events related to the core business are identified.

[0206] After identifying the target public opinion event, further analysis of the severity of the event is required, which includes evaluating the entities, relationships and scope of influence involved in the public opinion event and determining the potential impact of the event on the enterprise.

[0207] Analyze the impact of public opinion events on enterprises based on the entities and relationships in the public opinion events (such as cooperative relationships, competitive relationships, legal disputes, etc.). Quantify the severity of the events and classify the levels of the events, such as minor, moderate or severe, by analyzing the content of the public opinion events, the heat of public opinion and other indicators.

[0208] Depending on the severity of the incident, the system will generate a corresponding public opinion resolution strategy. Incidents of different severity require different responses to ensure that the company can respond to negative public opinion in a timely and effective manner.

[0209] Different public opinion resolution strategies are preset according to the severity of the incident (such as public relations statements, legal proceedings, adjustment of business strategies, etc.). The system automatically generates and recommends appropriate public opinion handling solutions for reference by corporate management.

[0210] Example description: In the financial industry, banks and securities companies need to keep abreast of negative public opinion information related to their core customers (such as large listed companies, debtors, etc.) in order to make risk management decisions. This example shows how to extract negative public opinion information based on the core business of the target entity, and analyze and process this information.

[0211] The public opinion monitoring system of financial institutions obtains public opinion data related to their core customers by crawling multiple information sources such as news websites, financial reports, social media, etc. The system forms a preliminary public opinion data set from these data sources, including positive, negative and neutral news reports of multiple companies.

[0212] The system pre-processes the collected public opinion data, removes irrelevant symbols and duplicate data, performs word segmentation, and processes the text into a format that is easy to analyze. For example, the system removes HTML tags and special symbols in news, and divides sentences into independent words or phrases to facilitate subsequent information extraction.

[0213] Financial institutions define the business entities of their core customers, such as company names, debtors, partners, etc. The system analyzes the entities in the collected public opinion data through the entity matching module to identify whether it contains content that matches the core business entities. For example, if a news report mentions a core customer of a financial institution and the report contains negative events such as "debt default" and "fund shortage", the event is identified as negative public opinion of the target subject.

[0214] For public opinion data related to core business entities, the system further uses sentiment analysis models to determine the positive or negative tendency of public opinion information. For example, if a report contains keywords such as "debt overdue" and "bankruptcy risk", the event is marked as a negative public opinion event. The system automatically filters out irrelevant or neutral news and only retains negative public opinion information related to the target entity.

[0215] After identifying negative public opinion, the system conducts a severity analysis based on the specific content of the event (such as whether it involves litigation or whether it has a significant financial impact). For example, the system conducts an in-depth analysis of news involving corporate debt defaults to determine whether the event will lead to an increase in the bank's loan risk. The system quantifies the impact of the event by setting a severity level for the event (such as "minor", "moderate", or "serious").

[0216] Based on the severity of the analysis, the system provides financial institutions with corresponding response strategies. For serious negative public opinion events, the system recommends that banks immediately conduct risk assessments, which may require adjustments to loan amounts and enhanced risk monitoring. For moderately serious events, the system recommends strengthening communication with customers and issuing announcements or clarification statements.

[0217] This embodiment effectively improves the accuracy of the enterprise's public opinion analysis by identifying core business entities and matching related public opinion events. Based on the matching analysis results, the system can automatically assess the severity of the public opinion event and generate corresponding public opinion resolution strategies to help enterprises quickly respond to potential risks. This solution improves the enterprise's risk management capabilities and reduces the negative impact of public opinion events.

[0218] In one embodiment, a negative public opinion information extraction device is provided, and the negative public opinion information extraction device corresponds one-to-one to the negative public opinion information extraction method in the above embodiment. Figure 3 , Figure 3 This is a functional module diagram of a preferred embodiment of the negative public opinion information extraction device of the present invention. Data acquisition module 10, data annotation module 20, model training module 30, framework construction module 40, data input module 50 and information extraction module 60. The functional modules are described in detail as follows:

[0219] The data acquisition module 10 is used to acquire public opinion information data related to the target subject, and form a public opinion information data set based on the public opinion information data;

[0220] A data annotation module 20 is used to define an information extraction structure according to information extraction requirements, annotate the public opinion information data set based on the information extraction structure, and generate a training data set, wherein the annotation includes annotation of entities, relationships, and events in the public opinion information data;

[0221] The model training module 30 is used to determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information;

[0222] A framework construction module 40, for constructing a processing framework for extracting negative public opinion information based on the information extraction model, wherein the processing framework includes an information extraction module;

[0223] A data input module 50 is used to obtain the public opinion data to be processed related to the target subject and input the public opinion data to be processed into the processing framework;

[0224] The information extraction module 60 is used for the information extraction module of the processing framework to identify the negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0225] In one embodiment, the data acquisition module 10 is specifically used for:

[0226] Performing text denoising processing on the public opinion information data;

[0227] Performing word segmentation processing on the processed public opinion information data to segment the text in the public opinion information data into single words or phrases;

[0228] Perform sentiment analysis on the text after word segmentation to determine the sentiment tendency of the text and label the text with the corresponding sentiment tendency label;

[0229] The public opinion information dataset is generated based on text annotated with sentiment tendency labels.

[0230] In one embodiment, the data annotation module 20 is specifically used to:

[0231] Defining an information extraction structure according to information extraction requirements, wherein the information extraction structure includes identification criteria for entities, relationships, and events in public opinion information data;

[0232] Based on the entity identification standard, the entities in the public opinion information dataset are labeled;

[0233] Based on the identification standard of the relationship, the relationship in the public opinion information dataset is labeled;

[0234] Based on the event identification criteria, marking the events in the public opinion information dataset;

[0235] The training data set is generated based on the annotated public opinion information data set.

[0236] In one embodiment, the model training module 30 is specifically used to:

[0237] Identify pre-trained models that have been trained on common text datasets;

[0238] Setting training parameters, and performing multiple rounds of iterative training on the pre-trained model based on the training parameters and the training data set;

[0239] In each round of training, the parameters of the pre-trained model are adjusted and optimized by a back-propagation algorithm;

[0240] After multiple rounds of iterative training are completed, the information extraction model for extracting negative public opinion information is obtained.

[0241] In one embodiment, the framework building module 40 is specifically used to:

[0242] A data input module of the configuration processing framework is used to receive public opinion information data to be processed;

[0243] A preprocessing module of the configuration processing framework is used to perform preprocessing operations on the input public opinion information data;

[0244] configuring the information extraction model as an information extraction module of a processing framework;

[0245] A feedback learning module of the configuration processing framework is used to analyze the difference between the current negative public opinion information and the historical negative public opinion data according to the extraction results of the information extraction module, determine the change pattern including event feature changes, sentiment tendency changes, entity relationship changes and event frequency changes, and store the change pattern as feedback data;

[0246] A dynamic update module of the configuration processing framework is used to analyze the feedback data stored in the feedback learning module and adjust the parameters of the information extraction model based on the analysis results.

[0247] In one embodiment, the information extraction module 60 is specifically used to:

[0248] Performing a preprocessing operation on the public opinion data to be processed by a preprocessing module of the processing framework;

[0249] Inputting the public opinion data to be processed after the preprocessing operation into the information extraction module;

[0250] The information extraction module analyzes the public opinion data to be processed according to the information extraction structure, and identifies negative public opinion information containing entity information, relationship information and event information in the public opinion data to be processed;

[0251] Generate target structured information based on the negative public opinion information including entity information, relationship information and event information.

[0252] In one embodiment, the information extraction module 60 is specifically used to:

[0253] Identify the core business entities related to the target entity;

[0254] The entity matching module is used to match and analyze entities in public opinion events and determine target public opinion events that match core business entities.

[0255] Analyze the severity of the target public opinion event based on the entities and relationships in the target public opinion event;

[0256] Generate a corresponding public opinion resolution strategy based on the severity of the target public opinion event.

[0257] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external user terminal through a network connection. When the computer program is executed by the processor, it realizes the functions or steps of a service-side method for extracting negative public opinion information.

[0258] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it realizes the functions or steps of a user-side method for extracting negative public opinion information

[0259] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following steps are implemented:

[0260] Acquire public opinion information data related to the target subject, and form a public opinion information data set based on the public opinion information data;

[0261] An information extraction structure is defined according to information extraction requirements, and the public opinion information data set is annotated based on the information extraction structure to generate a training data set, wherein the annotation includes annotation of entities, relationships and events in the public opinion information data;

[0262] Determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information;

[0263] Based on the information extraction model, a processing framework for extracting negative public opinion information is constructed, wherein the processing framework includes an information extraction module;

[0264] Acquire the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework;

[0265] The information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0266] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0267] Acquire public opinion information data related to the target subject, and form a public opinion information data set based on the public opinion information data;

[0268] An information extraction structure is defined according to information extraction requirements, and the public opinion information data set is annotated based on the information extraction structure to generate a training data set, wherein the annotation includes annotation of entities, relationships and events in the public opinion information data;

[0269] Determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information;

[0270] Based on the information extraction model, a processing framework for extracting negative public opinion information is constructed, wherein the processing framework includes an information extraction module;

[0271] Acquire the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework;

[0272] The information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

[0273] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant descriptions on the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0274] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0275] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0276] It should be noted that if software tools or components other than those of the Company appear in the embodiments of the present application, they are only used for illustration and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the above-mentioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the above-mentioned embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for extracting negative public opinion information, characterized in that: The following steps are involved: Acquire public opinion information data related to the target subject, and form a public opinion information data set based on the public opinion information data; An information extraction structure is defined according to information extraction requirements, and the public opinion information data set is annotated based on the information extraction structure to generate a training data set, wherein the annotation includes annotation of entities, relationships and events in the public opinion information data; Determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information; Based on the information extraction model, a processing framework for extracting negative public opinion information is constructed, wherein the processing framework includes an information extraction module; Acquire the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework; The information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

2. The method for extracting negative public opinion information according to claim 1, characterized in that: Acquiring public opinion information data related to the target subject, and forming a public opinion information data set based on the public opinion information data, including: Performing text denoising processing on the public opinion information data; Performing word segmentation processing on the processed public opinion information data to segment the text in the public opinion information data into single words or phrases; Perform sentiment analysis on the text after word segmentation to determine the sentiment tendency of the text and label the text with the corresponding sentiment tendency label; The public opinion information dataset is generated based on text annotated with sentiment tendency labels.

3. The method for extracting negative public opinion information according to claim 1, characterized in that: Defining an information extraction structure according to information extraction requirements, annotating the public opinion information dataset based on the information extraction structure, and generating a training dataset, including: Defining an information extraction structure according to information extraction requirements, wherein the information extraction structure includes identification criteria for entities, relationships, and events in public opinion information data; Based on the entity identification standard, the entities in the public opinion information dataset are labeled; Based on the identification standard of the relationship, the relationship in the public opinion information dataset is labeled; Based on the event identification criteria, marking the events in the public opinion information dataset; The training data set is generated based on the annotated public opinion information data set.

4. The method for extracting negative public opinion information as described in claim 1, characterized in that: Determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information, including: Identify pre-trained models that have been trained on common text datasets; Setting training parameters, and performing multiple rounds of iterative training on the pre-trained model based on the training parameters and the training data set; In each round of training, the parameters of the pre-trained model are adjusted and optimized by a back-propagation algorithm; After multiple rounds of iterative training are completed, the information extraction model for extracting negative public opinion information is obtained.

5. The method for extracting negative public opinion information according to claim 1, characterized in that: Based on the information extraction model, a processing framework for extracting negative public opinion information is constructed. The processing framework includes an information extraction module, including: A data input module of the configuration processing framework is used to receive public opinion information data to be processed; A preprocessing module of the configuration processing framework is used to perform preprocessing operations on the input public opinion information data; configuring the information extraction model as an information extraction module of a processing framework; A feedback learning module of the configuration processing framework is used to analyze the difference between the current negative public opinion information and the historical negative public opinion data according to the extraction results of the information extraction module, determine the change pattern including event feature changes, sentiment tendency changes, entity relationship changes and event frequency changes, and store the change pattern as feedback data; A dynamic update module of the configuration processing framework is used to analyze the feedback data stored in the feedback learning module and adjust the parameters of the information extraction model based on the analysis results.

6. The method for extracting negative public opinion information according to claim 1, characterized in that: The information extraction module of the processing framework identifies negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure, including: Performing a preprocessing operation on the public opinion data to be processed by a preprocessing module of the processing framework; Inputting the public opinion data to be processed after the preprocessing operation into the information extraction module; The information extraction module analyzes the public opinion data to be processed according to the information extraction structure, and identifies negative public opinion information containing entity information, relationship information and event information in the public opinion data to be processed; Generate target structured information based on the negative public opinion information including entity information, relationship information and event information.

7. The method for extracting negative public opinion information according to claim 1, characterized in that: After the information extraction module of the processing framework identifies the negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure, it also includes: Identify the core business entities related to the target entity; The entity matching module is used to match and analyze entities in public opinion events and determine target public opinion events that match core business entities. Analyze the severity of the target public opinion event based on the entities and relationships in the target public opinion event; Generate a corresponding public opinion resolution strategy based on the severity of the target public opinion event.

8. A negative public opinion information extraction device, characterized in that: The negative public opinion information extraction device comprises: A data acquisition module, used to acquire public opinion information data related to a target subject, and form a public opinion information data set based on the public opinion information data; A data annotation module, used to define an information extraction structure according to information extraction requirements, annotate the public opinion information data set based on the information extraction structure, and generate a training data set, wherein the annotation includes annotation of entities, relationships, and events in the public opinion information data; A model training module is used to determine a basic pre-training model, perform multiple rounds of iterative training on the basic pre-training model based on the training data set, update and optimize the parameters of the basic pre-training model, and generate an information extraction model for extracting negative public opinion information; A framework construction module, used to construct a processing framework for extracting negative public opinion information based on the information extraction model, wherein the processing framework includes an information extraction module; A data input module, used to obtain the public opinion data to be processed related to the target subject, and input the public opinion data to be processed into the processing framework; An information extraction module is used for the information extraction module of the processing framework to identify negative public opinion information including entities, relationships and events in the public opinion data to be processed according to the information extraction structure.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a negative public opinion information extraction program stored in the memory and executable on the processor. When the negative public opinion information extraction program is executed by the processor, the steps of the negative public opinion information extraction method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The storage medium stores a negative public opinion information extraction program, and when the negative public opinion information extraction program is executed by the processor, the steps of the negative public opinion information extraction method as described in any one of claims 1-7 are implemented.

Images