Illegal capital clue identification and early warning method and device

Abstract

The invention relates to the technical field of big data analysis, and particularly provides an illegal capital clue identification and early warning method and device, and the method comprises the following steps: S1, data collection and preprocessing; s2, text feature extraction and semantic modeling; s3, video content analysis and behavior recognition; s4, performing multi-modal fusion and risk scoring; and S5, early warning output and visual display are carried out. Compared with the prior art, the method has the advantages that potential illegal information can be automatically extracted from massive news reports and short video data, and efficient and accurate risk early warning is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and specifically provides a method and device for identifying and warning of illegal fundraising clues. Background Art

[0002] Illegal fundraising activities have become more diverse and covert in recent years. They attract public investment through false advertising and promises of high returns, severely disrupting financial market order and causing social instability. Traditional methods for identifying illegal fundraising rely primarily on manual judgment, regulatory reporting mechanisms, and statistical analysis of structured financial data. These methods suffer from delayed response times, limited coverage, and high rates of misjudgment.

[0003] With the development of the internet, illegal fundraising activities are increasingly using social media platforms to disseminate information, including news reports about investment projects that exaggerate returns and the production of misleading short video advertisements. These unstructured text and video content are difficult to effectively monitor and analyze using traditional methods.

[0004] Although some studies have attempted to use natural language processing (NLP) technology and image recognition algorithms to conduct preliminary screening of online information, due to limitations in semantic understanding capabilities and insufficient depth of multimodal fusion, the existing systems still have much room for improvement in recognition accuracy, real-time performance, and interpretability.

[0005] Therefore, there is an urgent need for an illegal fundraising clue identification and early warning method that combines large-scale language models with multimodal intelligent analysis, which can automatically extract potential illegal information from massive news reports and short video data to achieve efficient and accurate risk warnings. Summary of the Invention

[0006] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical method for identifying and warning of illegal fundraising clues.

[0007] A further technical task of the present invention is to provide a device for identifying and warning of illegal fundraising clues that is rationally designed, safe and applicable.

[0008] The technical solution adopted by the present invention to solve its technical problem is:

[0009] A method for identifying and warning of illegal fundraising clues has the following steps:

[0010] S1, data collection and preprocessing;

[0011] S2, text feature extraction and semantic modeling;

[0012] S3, video content analysis and behavior recognition;

[0013] S4, Multimodal Fusion and Risk Scoring;

[0014] S5. Early warning output and visual display.

[0015] Furthermore, in step S1, text data and video data from websites, platforms, and apps are collected, web pages are crawled to obtain text news data, and short videos are downloaded in batches using an API interface or simulated clicks;

[0016] During preprocessing, HTML tags, special characters, and stop words are removed, the encoding format is unified, key frame images are extracted at a fixed frame rate, OCR is performed on the text areas in the video, the video audio is converted into text content, duplicates are removed based on fields, and the data source category is marked.

[0017] Furthermore, in step S2, a large-scale language model is used to perform deep semantic analysis on the text, and a customized prompt template is used to guide the model to identify key words or sentence patterns related to illegal fundraising;

[0018] Output the risk keyword weight distribution, suspicious entity identification list, text risk score T for each text paragraph, and determine whether there is misleading propaganda intention based on the context.

[0019] Furthermore, in step S3, first, a convolutional neural network (CNN) is used to perform target detection and scene classification on the key frame image to identify whether it is an element disguised as a legitimate image, and the human behavior is detected through an action recognition model;

[0020] Then, after converting the audio into text, semantic analysis is performed based on the OCR results to extract key words and phrases related to illegal fundraising;

[0021] Finally, the time series model is used to analyze the character's movements; the overall risk score V of the video and the key frame / timestamp tags are output.

[0022] Furthermore, in step S4, first, the text semantic vector and the video authentication vector are spliced ​​and input into a unified large model for fusion. The text risk score T and the video risk score V are calculated respectively, and a comprehensive risk score F = α × T + β × V is generated through weighted averaging or attention mechanism, where α + β = 1;

[0023] Model distillation or adapter technology is used to integrate the knowledge of multiple LLM models into a unified framework to improve robustness.

[0024] Furthermore, after improving the robustness, risk scoring is performed. The risk score, keyword strength and entity matching degree of each modality are input, and a threshold θ is set. If F>θ, it is marked as a "high-risk clue".

[0025] Furthermore, in step S5, different levels of warning signals are generated according to the risk level, including low risk, medium risk and high risk, and a clue report is derived;

[0026] The map heat map shows high-risk areas; the risk trend map shows the frequency of illegal fundraising activities over time; the keyword cloud highlights high-frequency sensitive words; and it connects with the regulatory department system to realize automated reporting and disposal processes.

[0027] An illegal fund-raising clue identification and early warning device includes: at least one memory and at least one processor;

[0028] The at least one memory is configured to store a machine-readable program;

[0029] The at least one processor is used to call the machine-readable program to execute a method for identifying and warning of illegal fundraising clues.

[0030] Compared with the prior art, the method and device for identifying and warning of illegal fundraising clues of the present invention have the following outstanding beneficial effects:

[0031] This invention processes text news and short video content simultaneously, expands the monitoring scope, and combines LLM semantic understanding and multimodal fusion technology to significantly improve recognition accuracy. It supports minute-level updates and real-time warnings, provides keyword labeling, suspicious entity identification and specific risk descriptions, supports access to multiple language models and video parsing algorithms, and adapts to different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 It is a flowchart of a method for identifying clues of illegal fundraising and issuing early warnings. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0035] A best embodiment is given below:

[0036] like Figure 1 As shown, a method for identifying and warning of illegal fundraising clues in this embodiment has the following steps:

[0037] S1, data collection and preprocessing;

[0038] Collect text data and video data. The text data includes mainstream news websites (such as Caijing.com and Sina.com), social platforms (such as Weibo, Douyin, Kuaishou), and financial apps.

[0039] Video data includes short video platforms (such as Douyin, Kuaishou, and Bilibili), live streaming platforms, and corporate promotional videos.

[0040] Crawl web pages to obtain text news data, use API interfaces or simulated clicks to batch download short videos, and combine auxiliary information such as public reporting data, industrial and commercial registration data, public security, procuratorial and judicial data, and public opinion data.

[0041] During preprocessing, HTML tags, special characters, and stop words are removed, the encoding format is unified, key frame images are extracted at a fixed frame rate, OCR is performed on the text areas in the video, the video audio is converted into text content, duplicates are removed based on fields such as timestamps, IP addresses, account IDs, and the data source category is marked.

[0042] S2, text feature extraction and semantic modeling;

[0043] Use large-scale language models (such as Qwen, ChatGPT, BERT, etc.) to perform in-depth semantic analysis of text; customize prompt templates to guide the model to identify key words / sentence patterns related to illegal fundraising, such as "high returns", "guaranteed principal and interest", "risk-free returns", "internal resources", etc.; and typical language related to illegal fundraising, such as "short-term huge profits", "fund pool", "equity crowdfunding", etc.

[0044] Output the risk keyword weight distribution of each text paragraph, a suspicious entity identification list (such as company name, project name, contact information, etc.), and a text risk score (T). Combined with the context, determine whether there is misleading propaganda intent.

[0045] S3, video content analysis and behavior recognition;

[0046] Use convolutional neural networks (CNNs) to perform target detection and scene classification on keyframe images; identify whether they contain elements that disguise legitimate images, such as "office buildings," "conference sites," and "certificates and medals"; and use action recognition models (such as YOLOv8 and OpenPose) to detect human behaviors, such as "promoting products" and "distributing flyers."

[0047] After converting the audio into text, semantic analysis is performed in combination with the OCR results to extract key words and phrases related to illegal fundraising.

[0048] Use time series models (such as LSTM and Transformer) to analyze character movements; output the overall risk score (V) of the video and key frame / timestamp labels.

[0049] S4, Multimodal Fusion and Risk Scoring;

[0050] First, the text semantic vector and the video feature vector are concatenated and input into a unified large model for fusion. The text risk score (T) and video risk score (V) are calculated respectively. The comprehensive risk score F = α × T + β × V is generated through weighted averaging or attention mechanism, where α + β = 1.

[0051] Model distillation or adapter technology is used to integrate the knowledge of multiple LLM models into a unified framework to improve robustness.

[0052] Input the risk score, keyword strength, entity matching, etc. of each modality;

[0053] Set a threshold θ, if F>θ, it is marked as a “high-risk clue”.

[0054] S5, early warning output and visual display;

[0055] Generate different levels of warning signals (low risk, medium risk, high risk) according to the risk level; support exporting clue reports, including original content links, risk descriptions, suspicious entities, timestamp marks, etc.

[0056] The map heat map shows high-risk areas; the risk trend map shows the frequency of illegal fundraising activities over time; the keyword cloud highlights high-frequency sensitive words; and it supports docking with the regulatory system to realize automated reporting and disposal processes.

[0057] Based on the above method, an illegal fund-raising clue identification and early warning device in this embodiment includes: at least one memory and at least one processor;

[0058] The at least one memory is configured to store a machine-readable program;

[0059] The at least one processor is used to call the machine-readable program to execute a method for identifying and warning of illegal fundraising clues.

[0060] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0061] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and warning of illegal fundraising clues, characterized in that: The steps are as follows: S1, data collection and preprocessing; S2, text feature extraction and semantic modeling; S3, video content analysis and behavior recognition; S4, Multimodal Fusion and Risk Scoring; S5. Early warning output and visual display.

2. The method for identifying and warning of illegal fund-raising clues according to claim 1, characterized in that: In step S1, text data and video data from websites, platforms, and apps are collected, web pages are crawled to obtain text news data, and short videos are downloaded in batches using an API interface or simulated clicks; During preprocessing, HTML tags, special characters, and stop words are removed, the encoding format is unified, key frame images are extracted at a fixed frame rate, OCR is performed on the text areas in the video, the video audio is converted into text content, duplicates are removed based on fields, and the data source category is marked.

3. The method for identifying and warning of illegal fund-raising clues according to claim 2, characterized in that: In step S2, a large-scale language model is used to perform deep semantic analysis on the text, and a customized prompt template is used to guide the model to identify key words or sentence patterns related to illegal fundraising; Output the risk keyword weight distribution, suspicious entity identification list, text risk score T for each text paragraph, and determine whether there is misleading propaganda intention based on the context.

4. The method for identifying and warning of illegal fund-raising clues according to claim 3, characterized in that: In step S3, first, a convolutional neural network (CNN) is used to perform object detection and scene classification on the key frame image to identify whether it is an element disguised as a legitimate image, and the human behavior is detected through an action recognition model; Then, after converting the audio into text, semantic analysis is performed based on the OCR results to extract key words and phrases related to illegal fundraising; Finally, the time series model is used to analyze the character's movements; the overall risk score V of the video and the key frame / timestamp tags are output.

5. The method for identifying and warning of illegal fund-raising clues according to claim 4, characterized in that: In step S4, first, the text semantic vector and the video authentication vector are spliced ​​and input into a unified large model for fusion. The text risk score T and the video risk score V are calculated respectively. The comprehensive risk score F = α × T + β × V is generated through weighted average or attention mechanism, where α + β = 1; Model distillation or adapter technology is used to integrate the knowledge of multiple LLM models into a unified framework to improve robustness.

6. The method for identifying and warning of illegal fund-raising clues according to claim 5, characterized in that: After improving robustness, risk scoring is performed. The risk score, keyword strength, and entity matching degree of each modality are input, and a threshold θ is set. If F>θ, it is marked as a "high-risk clue." 7. The method for identifying and warning of illegal fund-raising clues according to claim 6, characterized in that: In step S5, different levels of warning signals are generated according to the risk level, including low risk, medium risk and high risk, and a clue report is derived; Map heatmaps show high-risk areas; The risk trend chart shows the frequency of illegal fundraising activities over time; Keyword cloud highlights high-frequency sensitive words; Connect with the regulatory department's system to realize automated reporting and disposal processes.

8. An illegal fund-raising clue identification and early warning device, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 7.

Images