Information management and analysis system based on image recognition and machine learning

Abstract

The invention discloses an information management and analysis system and method based on image recognition and machine learning. The system analyzes image data into a structured text through the image recognition module, performs text semantic analysis by using the natural language processing module, calculates quantitative indexes such as sensitive word scores and abnormal indexes in combination with sensitive word weight analysis, anomaly detection, a risk conduction map and a co-occurrence relation prediction model of the machine learning module, and performs semantic analysis on the structured text. And a risk index RI and an opportunity index OI are generated through a comprehensive decision module, and decision suggestions such as careful cooperation and preferential investment are output according to a preset rule. The model can dynamically optimize weight parameters based on historical data, and multi-modal data automatic processing and risk opportunity quantitative evaluation are achieved. According to the scheme, the decision making efficiency and scientificity are improved, the method is suitable for financial risk control, enterprise compliance and other scenes, and a traceable scientific basis is provided for commercial decision making.

Description

Technical Field

[0001] This article deals with information management and analysis systems based on image recognition and machine learning. Background Art

[0002] Traditional business information management and risk analysis relies primarily on manual review of text data such as policy documents, contracts, and emails, which suffers from the following flaws:

[0003] Inefficiency: Manually processing large amounts of text and image data is time-consuming and labor-intensive, especially due to the lack of automated parsing capabilities for image-based documents such as scanned contracts and invoices.

[0004] Highly subjective: Risk assessment relies on individual experience, and the scoring criteria are vague, which can easily lead to misjudgments or omissions;

[0005] Lack of dynamic adaptability: Traditional rule engines have difficulty coping with dynamic changes in complex business scenarios and are unable to automatically optimize analytical models.

[0006] With the diversification of commercial data (text, images, videos, etc.), there is an urgent need for an intelligent system that combines image recognition and machine learning to achieve automated processing and accurate analysis of multimodal data. Summary of the Invention

[0007] The purpose of this invention is to provide an information management and analysis system based on image recognition and machine learning. By automatically processing image and text data and combining it with machine learning models, it can achieve quantitative assessment of risks and opportunities, thereby improving the scientificity, efficiency, and accuracy of decision-making. The specific system is as follows:

[0008] An information management and analysis system based on image recognition and machine learning, including the following modules:

[0009] Image recognition module: Based on optical character recognition technology, it parses image data during the contact and transaction processes, including bidding documents, tender documents, contract documents, invoices, chat records, email correspondence and other materials, to obtain structured and unstructured text.

[0010] Natural language processing module: performs word segmentation, part-of-speech tagging, and named entity recognition on text data, and combines the existing keyword library to achieve semantic analysis of text content.

[0011] Machine Learning Module:

[0012] Build risk assessment and opportunity identification models, train sub-models based on historical data, and continuously improve the weights and parameter values ​​in the models.

[0013] The risk assessment and opportunity identification model includes a sensitive word weight analysis model, anomaly detection model, a risk transmission map model, and a co-occurrence relationship prediction model.

[0014] The sensitive word weight analysis model includes scoring sensitive words in policy documents, emails, instant chat records, and contracts during the transaction process. The specific scoring criteria are as follows:

[0015]

[0016] S ω : Total score of sensitive words;

[0017] ω i : The business weight of the i-th sensitive word;

[0018] TF_IDE(t i ): The importance of sensitive words in the current text;

[0019] Sentiment(t i ): Sentiment polarity value (risk words range from -1 to 0, opportunity words range from 0 to 1).

[0020] The anomaly detection model is based on statistical principles and identifies whether there are anomalies in structured data such as transaction amounts and transaction times. The specific calculation formula is as follows:

[0021]

[0022] E: abnormality index. When E is greater than or equal to 1, it is considered abnormal;

[0023] x: current data value;

[0024] μ: mean of historical data;

[0025] σ: standard deviation of historical data;

[0026] k: Industry correction coefficient. When the industry is a high-volatility industry, the k value is 1.5. When the industry is a stable industry, the k value is 0.8.

[0027] The risk transmission map model avoids the invisible risk transmission problem to calculate whether the new cooperative entity has a certain correlation with historical risks.

[0028]

[0029] R g : Graph associated risk value;

[0030] d ij : the shortest path length between new entity i and risk node j (number of cooperation steps);

[0031] RiskScore(n j ): Risk score of the historical risk node (0-100).

[0032] If the new partner company A and venture company B are connected through the "supplier-customer" relationship, then d 11 =2, B's historical risk score is 85, then

[0033] The co-occurrence relationship prediction model predicts this cooperation by verifying the degree of overlap between this cooperation and risk cooperation, that is, the co-occurrence relationship strength, to obtain the predicted direction of this cooperation. The specific calculation method is as follows:

[0034]

[0035] C ij : The co-occurrence relationship strength between entities i and j (cosine similarity value);

[0036] N i , N j : The set of associated nodes of entities i and j (including jointly participated projects, contracts, and tender documents).

[0037] After evaluating the above four sub-models, the integrated judgment of comprehensive risk assessment and opportunity identification is as follows:

[0038]

[0039] Among them, ∝, β, γ: risk dimension weights, the sum is 1;

[0040] α′, β′, γ′: opportunity dimension weights, the sum of which is 1;

[0041] S w ′: opportunity sensitive word score, at this time only the total score of positive sentiment words is calculated;

[0042] E′: trend value of the opportunistic structural characteristic. In this case, only the total score of positive opportunities is calculated.

[0043] Corresponding decision-making strategy:

[0044]

[0045] The above model can be used to manage and analyze information. The specific information management and analysis method includes the following steps:

[0046] S1, complete the analysis and semantic processing of multimodal data through the image recognition module and the NLP module;

[0047] S2. Determine and give the initial weights of each parameter based on historical data;

[0048] S3. Get the corresponding data and calculate S ω , E, R g 、C ij Numerical results of quantitative indicators;

[0049] S4. After eliminating negative data, calculate S w Numerical results of quantitative indicators of ′ and E′;

[0050] S5. Calculate the corresponding RI and OI indicators based on the comprehensive evaluation index calculation method, and generate cooperation suggestions according to the preset decision rules;

[0051] S6. If no suitable cooperation suggestions are obtained, the manual approval process will be entered, and cooperation suggestions will be made manually based on the data calculation process;

[0052] S71. If the manual suggestion is worthy of reference, the system will automatically record it and readjust the risk dimension weights ∝, β, γ and the opportunity dimension weights α′, β′, γ′ until the cooperation suggestion given by the system is consistent with the cooperation suggestion identified by the manual staff;

[0053] S72. If the manual suggestion is not worthy of reference, the system will automatically record it separately, and the review process will not be included in machine learning.

[0054] In order to further improve the adaptability of the model and the accuracy of calculations, a machine learning part is added to the model. The machine learning process is as follows:

[0055] S31. Obtain historical references and set initial risk dimension weights ∝, β, γ and opportunity dimension weights α′, β′, γ′;

[0056] S32. With the goal as the guide, the algorithm dynamically adjusts the risk dimension weights ∝, β, γ and the opportunity dimension weights α′, β′, γ′;

[0057] S33. When the set goal is reached, the weights of each parameter at this time are recorded and used as the initial historical reference for the next judgment or learning.

[0058] In order to better adapt to the cyclical changes in the industry and the market, the industry correction coefficient k in the anomaly detection model is dynamically generated based on industry fluctuation data.

[0059] At the same time, since many key business information is not obtained through public channels, key personnel must be given certain modification permissions. In particular, in the identification of related units, the shortest path length d between the new entity i and the risk node j is ijIt can be set to be generated by system calculation first, but humans have the highest modification authority, so that modifications can be made through real-time information acquisition to improve the accuracy of the final calculation.

[0060] Beneficial effects:

[0061] Compared with traditional manual experience judgment, this solution breaks down risk and opportunity assessment into quantifiable technical parameters. It uses mathematical formulas and algorithmic logic such as TF-IDF, cosine similarity, and outlier detection to clearly define the calculation rules for each parameter, avoiding subjective ambiguity and transforming business analysis from an "empirical art" to a "computable scientific process."

[0062] The effectiveness of this system relies on a robust data foundation, model design, and business scenario adaptability. It can enhance the transparency and reliability of decision-making and provide a scientific basis for human judgment. Furthermore, through parameterized definition and aggregated analysis, it can quickly and accurately identify risks and opportunities, providing decision-makers with guidance on how to make informed decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flowchart of an information management and analysis system based on image recognition and machine learning;

[0064] Figure 2 It is a distribution display diagram of decision-making strategies. DETAILED DESCRIPTION

[0065] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to embodiments and drawings. The embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0066] Example 1: Conduct analysis management training based on a completed contract.

[0067] S1. Set the result value requirements for the later stage, requiring the RI value to be between 50-80 and the 0I value to be greater than 40. The decision-making strategy is to proceed cautiously.

[0068] S2. First result prediction training:

[0069] S21. Image recognition processing: Scan the paper contract to generate an image file, extract the contract text, seal, amount capitalization and other information through the OCR module, and convert it into structured data;

[0070] Text semantic analysis: The NLP module detects sensitive words in the contract terms and calculates the sensitive word score S ω .

[0071] The customer's public opinion record contains risk words such as "breach of contract" and "responsibility", and the sentiment polarity value of such words is -0.8. At the same time, opportunity words such as the company's business record is rising and the technical advantage is high are also retrieved, and the sentiment polarity value of these words is 0.6. Then all the words are scored and summarized to obtain the final sensitive word score S ω Numeric value.

[0072] S22. Structured data verification: Extract the contract amount x = 50 million yuan, compare it with the historical mean μ = 30 million yuan and the standard deviation σ = 8 million yuan, and based on the company's stable industry, set K to 0.8. Calculate the anomaly index E = 2 and determine it as abnormal;

[0073] S23, Risk Transmission Analysis: New partner company A connects to risk company B through a two-step partnership. ij =2, B's risk score is 85, calculate

[0074] S24. The machine learning module automatically sets the risk dimension weights to ∝ = 0.4, β = 0.3, and γ = 0.3 based on historical similar contract data, and calculates the comprehensive risk index RI = 37.35;

[0075] S25. At this time, opportunity identification and judgment are performed again, and the opportunity dimension weights α′=0.4, β′=0.3, and γ′=0.3 are set, and the comprehensive opportunity index OI=42.35 is calculated;

[0076] S26. Compared with the ideal output value, the output value at this time is too high for risk assessment, but relatively ideal for opportunity assessment. Therefore, the risk assessment process needs to be further optimized and learned.

[0077] S27. The machine learning module automatically adjusts the risk dimension based on the results generated in the later stage of this cooperation until the RI value in the output result is between 50-80.

[0078] S3. At this time, the weights of each dimension of this type of contract are recorded. The next time there is a similar cooperation, the system can refer to the experience of this cooperation, thereby achieving mechanical self-learning of the system, thereby improving the judgment of similar behaviors next time.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An information management and analysis system based on image recognition and machine learning, characterized by: include: An image recognition module configured to parse image data using OCR technology to extract structured text and unstructured text; A natural language processing module configured to perform word segmentation, named entity recognition, and sensitive word sentiment analysis on text data; The machine learning module includes four sub-models: a sensitive word weight analysis model, an anomaly detection model, a risk transmission map model, and a co-occurrence relationship prediction model. The results of each sub-model are calculated and input into the comprehensive decision-making module. The comprehensive decision-making module generates decision recommendations based on the risk index RI and the opportunity index OI. The integrated judgment of comprehensive risk assessment and opportunity identification is as follows: Among them, ∝, β, γ: risk dimension weights, the sum is 1; α′, β′, γ′: opportunity dimension weights, the sum of which is 1; S w : Total score of sensitive words; E: abnormality index; R g : Graph associated risk value; S w ′: opportunity-sensitive word score, at this time only the total score of positive sentiment words is calculated; E′: trend value of the opportunistic structural characteristic. In this case, only the total score of positive opportunities is calculated. Corresponding decision-making strategy:

2. The information management and analysis system based on image recognition and machine learning according to claim 1, characterized in that: Sentiment weight analysis model: S ω : Total score of sensitive words; ω i : The business weight of the i-th sensitive word; TF_IDE(t i ): The importance of sensitive words in the current text; Sentiment(t i ): Sentiment polarity value (risk words range from -1 to 0, opportunity words range from 0 to 1); Anomaly Detection Model: E: abnormality index. When E is greater than or equal to 1, it is considered abnormal; x: current data value; μ: mean of historical data; σ: standard deviation of historical data; k: Industry correction factor.

3. The information management and analysis system based on image recognition and machine learning according to claim 2, characterized in that: Risk transmission map model: R g : Graph associated risk value; d ij : the shortest path length between new entity i and risk node j; RiskScore(n j ): Risk score of historical risk node.

4. The information management and analysis system based on image recognition and machine learning according to claim 2 or 3, characterized in that: Co-occurrence relationship prediction model: C ij : the co-occurrence relationship strength between entities i and j; N i , N j : The set of associated nodes of entity i, j.

5. The information management and analysis system based on image recognition and machine learning according to claim 4, characterized in that: The process of machine learning is: S31. Obtain historical references and set initial risk dimension weights ∝, β, γ and opportunity dimension weights α′, β′, γ′; S32. With the goal as the guide, the algorithm dynamically adjusts the risk dimension weights ∝, β, γ and the opportunity dimension weights α′, β′, γ′; S33. When the set goal is reached, the weights of each parameter at this time are recorded and used as the initial historical reference for the next judgment or learning.

6. The information management and analysis system based on image recognition and machine learning according to claim 4, characterized in that: The industry correction coefficient k in the anomaly detection model is dynamically generated based on industry fluctuation data.

7. The information management and analysis system based on image recognition and machine learning according to claim 4, characterized in that: The shortest path length d between new entity i and risk node j ij Generated through system calculation, humans have the highest modification authority.

8. An information management and analysis method, characterized in that: The following steps are involved: S1, complete the analysis and semantic processing of multimodal data through the image recognition module and the NLP module; S2. Determine and give the initial weights of each parameter based on historical data; S3. Get the corresponding data and calculate S ω , E, R g 、C ij Numerical results of quantitative indicators; S4. After eliminating negative data, calculate S w Numerical results of quantitative indicators of ′ and E′; S5. Calculate the corresponding RI and OI indicators based on the comprehensive evaluation index calculation method, and generate cooperation suggestions according to the preset decision rules; S6. If no suitable cooperation suggestions are obtained, the manual approval process will be entered, and cooperation suggestions will be made manually based on the data calculation process; S71. If the manual suggestion is worthy of reference, the system will automatically record it and readjust the risk dimension weights ∝, β, γ and the opportunity dimension weights α′, β′, γ′ until the cooperation suggestion given by the system is consistent with the cooperation suggestion identified by the manual staff; S72. If the manual suggestion is not worthy of reference, the system will automatically record it separately, and the review process will not be included in machine learning.

Images