Generative AI-based risk report analysis method and equipment for department and creative enterprises, and medium

Abstract

The invention provides a generative AI-based risk report analysis method and equipment for a department and wound enterprise and a medium, belongs to the technical field of enterprise risk report analysis, and aims to identify the field of department and wound to which the enterprise belongs, obtain corresponding risk configuration and define a modular report framework and quantitative indexes of all dimensions. Multi-source heterogeneous data are integrated through an ETL process, and multi-modal features are extracted and fused. And performing risk prediction by utilizing a generative AI model, outputting scores and overall levels of all dimensions, generating a professional report by adopting a retrieval enhancement generation technology in combination with an evidence library, and performing visual display. According to the method, the full-process automatic generation of the risk report is realized, and the efficiency is greatly improved. And through multi-modal data fusion and field adaptive configuration, the analysis depth and accuracy are ensured. And a retrieval enhancement generation framework is adopted, so that the content authenticity is effectively guaranteed, and the risk assessment requirements of different department and wound fields can be quickly met.

Description

Technical Field

[0001] This invention belongs to the field of enterprise risk report analysis technology, specifically involving a method, equipment, and medium for analyzing risk reports of science and technology innovation enterprises based on generative AI. Background Technology

[0002] Risk assessment for science and technology innovation enterprises is an important application area where FinTech, corporate credit investigation, investment research and analysis, and RegTech converge. The rapid development of generative artificial intelligence (GAI) technology provides a new path for automated and intelligent risk analysis.

[0003] In related technologies, financial indicators are used to analyze corporate risks, but the risks specific to the science and technology innovation field, such as the risk of clinical trial failure in biomedicine and the risk of chip R&D cycle in semiconductors, are ignored, resulting in inaccurate risk identification.

[0004] Risk assessment involves data processing, metric calculation, and report writing. Current methods are inefficient and influenced by past experience; the same company's risk level may be assessed as medium or high by different analysts, lacking objective standards. While AI-powered report generation technology is used, the model-generated text lacks integration with actual data evidence, leading to reports containing content inconsistent with the company's actual situation and reducing their credibility. Summary of the Invention

[0005] This invention provides a risk report analysis method for science and technology innovation enterprises based on generative AI, which improves the efficiency and accuracy of risk assessment, enables in-depth mining of multimodal data, and enhances the professionalism and credibility of the report.

[0006] The methods include: S101: Identify the science and technology innovation field to which the enterprise belongs and obtain the risk characteristics configuration of the science and technology innovation field; S102: Define a modular reporting framework, including multiple risk dimensions, and configure a data requirements list and feature extraction rules for each dimension; S103: For each risk dimension in the reporting framework, based on the risk characteristics configuration obtained in S101, define quantitative and qualitative indicators; S104: Based on the data requirement list in S102 and the indicators defined in S103, acquire enterprise data from the database and construct a multi-source heterogeneous dataset through ETL. S105: Based on the feature extraction rules of S102 and the metrics defined in S103, extract text, image and time series data features from multi-source heterogeneous datasets, and perform feature fusion to obtain fused multimodal features; S106: Input the fused multimodal features obtained in S105 into a generative AI model for enterprise risk prediction tasks, and output the risk scores of each risk dimension and the overall risk level. S107: Based on the risk score and risk level obtained in S106, and combined with relevant evidence retrieved from the knowledge base, a qualitative analysis paragraph is generated using a text generation model based on a retrieval-enhanced generation architecture, and assembled into a complete risk report according to the report framework defined in S102. S108: Visualize the complete risk report generated in S107.

[0007] Preferably, S102 specifically includes the following methods: Establish a tree-structured risk dimension framework that includes multiple primary risk dimensions and their subordinate secondary sub-dimensions; For each secondary sub-dimension in the risk dimension framework, define a standardized list of data requirements, including data type, data source, and data fields; Based on the data requirements list, configure text feature extraction, image feature extraction, and time series data feature extraction methods for each secondary sub-dimension; Construct a data association mapping table and a feature interaction mapping table between the secondary sub-dimensions in the risk dimension framework.

[0008] Preferably, S103 specifically includes the following methods: Call the risk feature configuration output by S101 and load the set of core indicator items corresponding to the current science and technology innovation field; Based on the risk dimensions in the reporting framework defined in S102, at least one quantitative indicator and at least one qualitative indicator are selected and bound for each dimension from the set of core indicator items. Configure the required data source fields and calculation formula expressions for each bound quantitative indicator; Configure the data sources and key identification tags required for analysis for each bound qualitative indicator; The mapping relationships between all indicators and data sources and rules are integrated to generate indicator configuration files.

[0009] Preferably, S105 specifically includes the following methods: Based on the feature extraction rules described in S102 and the metrics defined in S103, different feature extractors are scheduled from the multi-source heterogeneous dataset to extract features from text, images, and time-series data respectively. The extracted text features are parsed and vectorized to generate text feature vectors. The extracted image features and time-series data features are standardized and serialized respectively to generate image feature vectors and time-series feature vectors; Text feature vectors, image feature vectors, and temporal feature vectors are projected and aligned based on a pre-defined unified dimensional space, and then concatenated to form a fused multimodal feature vector. The fused multimodal feature vectors are associated with the enterprise identifier.

[0010] Preferably, S106 specifically includes the following methods: The fused multimodal feature vectors output by S105 are standardized and preprocessed to map the feature values ​​to the input numerical range of the generative AI risk prediction model, thus adapting to the input dimension and distribution requirements of the model. The generative AI risk prediction model, which has been pre-optimized by domain adaptation and reinforcement learning, is invoked. The generative AI risk prediction model takes standardized multimodal features as input. During model runtime, the risk dimension association module dynamically weights the encoded features using a predefined risk dimension-feature weight matrix; Based on the weighted feature vector, the output generation module calculates the continuous risk score for each risk dimension through a linear regression layer and outputs the risk level through a Softmax classification layer. The output scores for each risk dimension and the overall level are sorted according to the risk dimension order defined in the reporting framework of S102 to generate risk scores and risk levels.

[0011] Preferably, S1062 specifically includes the following methods: A training corpus for risk analysis of science and technology innovation enterprises was constructed, and domain risk feature configurations and multi-source heterogeneous datasets were loaded as basic training samples. Based on the general generative AI foundation model, the training corpus is used for domain-adaptive pre-training, and the model parameters are updated to adapt to the semantic features of the risk domain of science and technology innovation enterprises. Based on the pre-trained model, supervised fine-tuning is performed using a sample dataset labeled with scores for each risk dimension and the overall risk level to optimize the model's risk identification accuracy.

[0012] Preferably, S107 specifically includes the following methods: Read the risk assessment result table output by S106 and the dimension attribute table containing risk dimension names and core analysis objectives in the modular report framework of S102, and generate an input list by associating them through risk dimension IDs; Based on the risk dimension ID in the generated input list and the indicator ID in the indicator configuration package, retrieve evidence data related to each dimension and indicator; Load the text generation model file based on the retrieval-enhanced generation architecture, read the qualitative description specifications in the S102 reporting framework and the domain expression style in the S101 risk feature configuration, and write the specification and style parameters into the model's generation constraint module through the model configuration tool; The dimension names, risk scores, and evidence data from the corresponding evidence folder in the generated input list are combined into the model input text according to the preset format. The configured text generation model generates qualitative analysis paragraphs for each risk dimension.

[0013] Preferably, S1073 specifically includes the following methods: Load the pre-trained text generation model file; Read qualitative description specification data from the report framework configuration; Read the domain representation style parameters from the risk feature configuration, including a list of domain-specific terms and representation style rules; The qualitative description specifications and domain representation style parameters are parsed by the model configuration tool and converted into a configuration format that the model generation constraint module can recognize. Write the converted configuration parameters into the model's generation constraint module to update the generation parameter settings of the model instance.

[0014] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the generative AI-based risk report analysis method for science and technology innovation enterprises.

[0015] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the generative AI-based risk report analysis method for science and technology innovation enterprises.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a generative AI-based risk report analysis method for science and technology innovation enterprises. By identifying the enterprise's science and technology innovation field and acquiring its unique risk feature configuration, and defining indicators based on the field configuration, the risk analysis focuses on the core risks within that field, avoiding biases caused by general analysis. The defined data requirements list and feature extraction rules guide the ETL process, making the collection, cleaning, and transformation of text, image, and time-series data more targeted, reducing redundant data, and improving data utilization efficiency. The integration of text, image, and time-series features compensates for the information limitations of different data types. The generative AI model, based on multimodal feature prediction, reduces the subjectivity of the assessment and improves the quantitative accuracy of risk scores and levels. Combining knowledge base evidence to generate qualitative analysis reduces the mismatch between the generative model and the actual situation of the enterprise, and the visualization transforms complex risk information into intuitive charts, facilitating rapid understanding of core risks. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the risk report analysis methodology for science and technology innovation enterprises based on generative AI; Figure 2 This is a schematic diagram of an electronic device. Detailed Implementation

[0019] This invention provides a generative AI-based risk report analysis method for science and technology innovation enterprises. By constructing a multi-source data real-time acquisition and stream processing pipeline, it integrates real-time signals from publicly available online information, government data interfaces, news and public opinion, and industry chain dynamics. Utilizing the temporal understanding capabilities of a large-scale model, it captures immediate changes affecting the enterprise's risk status, improving the timeliness and foresight of the risk report. It can deeply fuse and jointly infer multimodal data such as structured financial data, patents, research reports, bidding documents, and correlation graphs. Through the powerful semantic understanding and generation capabilities of the Large Language Model (LLM), the system can extract key risk signals and uncover implicit relationships between data, thereby constructing a comprehensive and three-dimensional enterprise profile. This invention leverages the capabilities of generative large-scale models to transform the analysis results into a logically clear and well-supported narrative risk report. Each conclusion in the report is automatically linked to the analysis basis and data source and visually displayed using embedded charts, enhancing the transparency of the risk assessment process and the credibility of the results.

[0020] The overall architecture design of this invention can be based on the following approach: The system should adopt a layered and decoupled microservice architecture to ensure flexibility, scalability, and maintainability.

[0021] (a) User Interaction Layer The web front-end or API interface receives user queries and displays the generated reports in a visual web page, supporting downloads in PDF and Word formats.

[0022] (ii) Application Service Layer Task scheduling engine: Coordinates the entire report generation process, sequentially calling services such as data acquisition, model analysis, and report assembly. Report generation and assembly service: Populates the analysis results from each module into the underlying report framework template to generate the final report.

[0023] (III) AI Model Layer Multimodal feature extraction model: processes data of different modalities such as text, data, and charts.

[0024] Multimodal fusion and reasoning model: integrates all features for joint reasoning and risk assessment.

[0025] Text generation model: Based on structured conclusions, it generates fluent and professional qualitative analysis paragraphs.

[0026] (iv) Data layer: Multi-source data acquisition and access: Integrating internal and external data sources.

[0027] Data lake / data warehouse: Stores raw data and processed data.

[0028] Vector database: Stores embedded vectors such as news and announcements for efficient similarity retrieval.

[0029] Infrastructure layer: cloud platform or local server, providing computing, storage and network resources.

[0030] The following describes in detail the generative AI-based risk report analysis method for science and technology innovation enterprises involved in this application. Specific details, such as particular system structures and technologies, are presented for illustrative purposes rather than limiting, to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0031] It should be understood that, when used in this specification, terms include indicating the presence of a described feature, integral, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms include, encompass, have, and variations thereof mean including but not limited to, unless otherwise specifically emphasized.

[0032] The statements such as "one embodiment" or "some embodiments" described in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the statements such as "in one embodiment," "in some embodiments," "in other embodiments," and "in still other embodiments" in this application do not necessarily refer to the same embodiment, but rather mean one or more, but not all, embodiments, unless otherwise specifically emphasized.

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figure 1 The diagram shows a flowchart of a risk report analysis method for science and technology innovation enterprises based on generative AI in a specific embodiment. The method includes: S101: Identify the science and technology innovation field to which the enterprise belongs and obtain the risk characteristics configuration of the science and technology innovation field.

[0035] In some embodiments, basic definition standards are set, combined with official science and technology innovation qualification certifications, including science and technology-based SMEs, innovative SMEs, high-tech enterprises, specialized and innovative SMEs, specialized and innovative little giants, private science and technology enterprises, enterprise technology centers, engineering technology research centers, hidden champion enterprises, hidden champion demonstration enterprises, technology innovation demonstration enterprises, fledgling enterprises, unicorn enterprises, gazelle enterprises, future unicorn enterprises, science and technology business incubators, technology-advanced service enterprises, technology innovation demonstration enterprises, innovative enterprises, innovative leading enterprises, manufacturing single champion enterprises, manufacturing single champion demonstration enterprises, manufacturing single champion cultivation enterprises, seed unicorn enterprises, makerspaces, potential unicorn enterprises, potential unicorn enterprises, intellectual property advantage enterprises, intellectual property demonstration cultivation enterprises, intellectual property standardization enterprises, intellectual property demonstration enterprises, intellectual property advantage cultivation enterprises, key laboratories, etc., as well as the listing standards for the Science and Technology Innovation Board.

[0036] Using NLP technology, keyword recognition and industry classification are performed. Based on company name, business scope, patent information, product introduction, etc., combined with the China Securities Regulatory Commission's industry classification and strategic emerging industries classification standards, the system can automatically determine whether a company belongs to a science and technology innovation enterprise and its sub-sector.

[0037] It should be noted that the intelligent risk report for science and technology innovation enterprises needs to be modularly designed, that is, the report should be divided into multiple chapters, each chapter corresponding to a risk dimension, with a configurable template engine, chapter-level JSON Schema definition, and drag-and-drop template arrangement for technical implementation.

[0038] Furthermore, regarding quantitative and qualitative indicators, quantitative indicators include financial indicators such as debt-to-equity ratio, gross profit margin, R&D expense ratio, and revenue compound annual growth rate; innovation indicators such as the number of patent applications / grants, patent citations, and the proportion of R&D personnel; and market indicators such as market share, customer concentration, and supplier concentration. Qualitative indicators include technological advancement: textual analysis of patent abstracts and specifications, and comparison with industry technical terminology databases; team background: education of the senior management team, work experience extracted from annual reports and prospectuses; and competitive position: descriptions of the company in press releases and industry reports, such as industry leader, pioneer, or innovator.

[0039] This embodiment uses the company's main business keywords, patent classification numbers, and industry qualification certifications from its registration information to compare them with a preset list of science and technology innovation fields using a field matching tool to determine the field to which the company belongs. It then retrieves a risk characteristic configuration file containing field-specific risk types.

[0040] S102: Define a modular reporting framework that includes multiple risk dimensions and configure a list of data requirements and feature extraction rules for each dimension.

[0041] In some embodiments, a template editor is used to define the module hierarchy of the report framework, with each module level corresponding to a risk dimension, such as financial risk, technology and innovation risk, and market risk. A data requirement list is configured for each dimension, allowing users to record the source identifiers, field names, and update frequencies of the required data using a table tool. For example, financial risk corresponds to fields such as debt-to-equity ratio and net cash flow. Rule templates are used to define the keyword range for text features, the extraction area for image features, and the time window for time-series data. For example, text features for technology and innovation risk require extracting keywords such as patent authorizations and the number of R&D personnel. This ensures clear analytical logic and a well-defined hierarchy, improving process efficiency.

[0042] S102 specifically includes the following methods: Step S1021: Based on the risk characteristics of science and technology innovation enterprises, establish a multi-level risk dimension classification system, including primary dimensions such as industry and market risk, technology and innovation risk, and financial and operational risk. Each primary dimension has 3-5 secondary sub-dimensions, forming a tree-structured risk dimension framework.

[0043] Step S1022: For each secondary risk dimension, define a standardized data requirements list template, including elements such as data type, data source, data fields, and data frequency.

[0044] This embodiment designs a standardized data requirements checklist template, translating each risk dimension into specific data collection requirements, clearly defining the data types, sources, and formats to be acquired. This guides the data collection process, ensuring the integrity and standardization of the acquired data.

[0045] Step S1023: Develop a feature extraction rule base for each secondary risk dimension, including text feature extraction rules, image feature extraction rules, and price fluctuation trends extracted from market data.

[0046] This embodiment establishes a feature extraction rule base for different data types, transforming raw data into analyzable feature vectors. It supports feature extraction from text, images, and time-series data, ensuring the comparability of features from different sources and providing high-quality feature input for risk assessment.

[0047] Step S1024: Construct a correlation mapping table between risk dimensions to clarify the data dependencies and feature interactions between each dimension. For example, there is a positive correlation between the technological advancement feature in the technology and innovation risk dimension and the R&D investment ratio feature in the financial and operational risk dimension.

[0048] Step S1025: Define a visual configuration management interface that allows users to adjust the risk dimension structure, modify data requirement list parameters, and update feature extraction rules by dragging and dropping, and generate configuration results in a file format.

[0049] S103: For each risk dimension in the reporting framework, define quantitative and qualitative indicators based on the risk characteristics configuration obtained in S101.

[0050] In some embodiments, the domain risk characteristics configured in S101 are associated with the risk dimensions in S102. Quantitative indicators are defined for quantifiable characteristics. Qualitative indicators are defined for characteristics requiring descriptive evaluation. The indicator definitions are tied to domain characteristics to ensure that the evaluation criteria conform to industry characteristics and improve the professionalism of the results.

[0051] S104: Based on the data requirement list in S102 and the indicators defined in S103, obtain enterprise data from the database and construct a multi-source heterogeneous dataset through ETL.

[0052] In some embodiments, multi-source databases are connected based on the source identifiers in the data requirement list in S102. ETL tools are used to perform data extraction, cleaning, and transformation, and the processed data is finally classified by enterprise ID and risk dimension and stored as a multi-source heterogeneous dataset.

[0053] S105: Based on the feature extraction rules described in S102 and the indicators defined in S103, extract text, image, and time-series data features from the multi-source heterogeneous dataset, and perform feature fusion to obtain fused multimodal features.

[0054] In some embodiments, different types of data are processed according to the rules in S102: keyword frequency is extracted from text data; numerical information and technical identifiers are identified from image data; trend values ​​and fluctuation ranges are calculated for monthly R&D investment. Then, text, image, and time-series features under the same risk dimension are combined by indicator ID to form a fused multimodal feature vector. Features are bound to risk dimensions and indicators to ensure that the features input into subsequent models directly serve risk prediction and improve prediction relevance.

[0055] S106: Input the fused multimodal features obtained in S105 into a generative AI model for enterprise risk prediction tasks, and output the risk scores of each risk dimension and the overall risk level.

[0056] In some embodiments, the fused multimodal feature vectors are standardized, such as mapped to the 0-100 range, to adapt to the input format of the generative AI model. The model loads pre-trained risk prediction parameters and fine-tunes them using domain data to process the input features. The processing involves converting multimodal features into a unified semantic vector through a feature encoding module, strengthening the weights of domain-related features through a risk dimension association module, and finally calculating the scores for each risk dimension through an output generation module. These scores are then aggregated according to preset weights to obtain the overall risk level, which can be categorized as low / medium / high.

[0057] S107: Based on the risk score and risk level obtained in S106, and combined with relevant evidence retrieved from the knowledge base, a qualitative analysis paragraph is generated using a text generation model based on a retrieval-enhanced generation architecture, and assembled into a complete risk report according to the report framework defined in S102.

[0058] In some embodiments, relevant evidence is retrieved using the risk dimension ID (S106) and indicator ID (S103) as keywords. A text generation model based on the RAG architecture is invoked, inputting risk scores, evidence fragments, and the report framework format requirements (S102) to generate qualitative analysis paragraphs. These paragraphs are then assembled according to the framework order (S102) to form a complete report, improving its reliability; assembling the report according to a preset framework ensures a standardized report structure, facilitating reading and use.

[0059] S108: Display the complete risk report generated in S107 in a visual manner.

[0060] In one embodiment of the present invention, based on step S103, the following is a possible embodiment and its specific implementation is described in a non-limiting manner. S103 specifically includes the following methods: S1031: Call the risk feature configuration output by S101 and load the set of core indicator items corresponding to the current science and technology innovation field; S1032: Based on the risk dimensions in the reporting framework defined in S102, select and bind at least one quantitative indicator and at least one qualitative indicator for each dimension from the set of core indicator items; S1033: Configure the required data source fields and calculation formula expressions for each bound quantitative indicator; In this embodiment, the calculation formula expression can be an expression stored in string form, such as R&D expenses / operating revenue * 100%.

[0061] S1034: For each bound qualitative indicator, configure the data source and key identification identifier required for analysis; S1035: Integrate the mapping relationships between all indicators and data sources and rules to generate indicator configuration files.

[0062] For example, taking the biopharmaceutical field as an example, S1031: Call the risk feature configuration output by S101 and load the set of core indicator items corresponding to the current science and technology innovation field. Combined with the risk feature configuration of the biopharmaceutical field output by S101, it includes the set of indicator items for R&D risk, compliance risk, and production risk in this field.

[0063] Optionally, it may include the following: Quantitative indicators: clinical trial progress completion rate, R&D expenditure ratio, production qualification rate, and regulatory compliance inspection pass rate; Qualitative indicators include: stability of the core R&D team, scientific validity of clinical trial protocols, maturity of manufacturing processes, and completeness of drug registration application materials.

[0064] S1032: Based on the risk dimensions in the reporting framework defined in S102, at least one quantitative indicator and at least one qualitative indicator are selected and bound to each dimension from the set of core indicator items. The reporting framework defined in S102 includes three risk dimensions: R&D risk, compliance risk, and production risk. The binding results are as follows: R&D risks: Quantitative indicators are tied to the completion rate of clinical trials, while qualitative indicators are tied to the stability of the core R&D team; Production risk: Quantitative indicators are tied to the production qualification rate, while qualitative indicators are tied to the maturity of the production process.

[0065] S1033: Configure the required data source fields and calculation formula expressions for each bound quantitative indicator.

[0066] Optionally, the clinical trial progress completion rate is calculated using data from the actual completed stages and the total planned stages in the clinical trial management system. The calculation formula is stored as: (Number of actual completed stages / Total planned stages) * 100%. Production pass rate: The data source is the qualified product quantity field and the total production quantity field of the production record database; the calculation formula is stored as qualified product quantity / total production quantity * 100%.

[0067] S1034: Configure the data sources and key identification tags required for analysis for each bound qualitative indicator.

[0068] Stability of the core R&D team: Data sources are the R&D team section of the company's annual report and the core member departure records of the human resources system; key identification indicators are that the number of core members (with doctoral degrees or above and who participated in core projects) leaving the company is ≤1 person / year and the average tenure of the team is ≥5 years.

[0069] Completeness of drug registration application materials: Data source is the drug registration application system - material submission record; key identification indicators are missing material items = 0 and the number of times materials were supplemented ≤ 1.

[0070] Production process maturity: Data sources are production process documents and process validation reports; key identification indicators are process parameter fluctuation range ≤ ±5% and three consecutive batches of trial production meeting the standards.

[0071] S1035: Integrate the mapping relationships between all indicators and data sources and rules to generate an indicator configuration file. The integrated indicator configuration file provides unified guidance for data collection and feature extraction, reduces information bias across steps, and improves the overall process efficiency and the accuracy of risk analysis.

[0072] In one embodiment of the present invention, based on step S105, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S105 specifically includes the following: S1051: Based on the feature extraction rules described in S102 and the metrics defined in S103, different feature extractors are scheduled from the multi-source heterogeneous dataset to extract features from text, images, and time-series data respectively.

[0073] S1052: Perform entity parsing and vectorization on the extracted text features to generate text feature vectors.

[0074] S1053: The extracted image features and time-series data features are standardized and serialized respectively to generate image feature vectors and time-series feature vectors.

[0075] S1054: Project and align the text feature vector, image feature vector, and temporal feature vector based on a preset unified dimensional space, and then concatenate them to form a fused multimodal feature vector.

[0076] S1055: Associate the fused multimodal feature vector with the enterprise identifier and store it in the feature library.

[0077] The following is a practical example using a biopharmaceutical technology innovation company, A. S101 identifies company A's field as biopharmaceutical. S102 defines risk dimensions including technology and innovation risk, with data requirements including: patent texts, clinical trial announcements (text), scanned financial statements (images), and R&D investment data (time series) for the past three years. Feature extraction rules require: extracting technical keywords from patents, identifying R&D expense figures from financial statement images, and extracting the R&D investment time series from annual reports. S103 defines indicators including: a qualitative indicator of core technology advancement, and a quantitative indicator of R&D expense ratio.

[0078] S1051 schedules different feature extractors for multimodal feature extraction, scheduling three types of extractors according to the rules in S102 and the metrics in S103: The text extractor is invoked to process Company A's patent abstracts and clinical trial termination notices; Use an image extractor to process a screenshot of the profit and loss statement page from Company A's 2023 financial report PDF; The time series extractor is invoked to read the R&D expenses and operating revenue fields of company A from 2021 to 2023 from the structured database.

[0079] The text extractor in S1052 identifies technical entities from the patent abstract: mRNA delivery system and lipid nanoparticles; extracts events from clinical trial announcements: Phase III clinical trial terminated early due to insufficient efficacy; inputs the above text into a pre-trained language model encoder to generate a 768-dimensional text feature vector T, which integrates technical keywords and negative event semantics.

[0080] The standardization and serialization of image and temporal features in S1053 is based on the image extractor recognizing R&D expenses and operating income in financial statement images through OCR, calculating the R&D expense ratio, converting it into a scalar and expanding it into a fixed-length vector, and then normalizing it to obtain the image feature vector.

[0081] The temporal extractor acquires three years of R&D investment data, aligns the timestamps, performs Z-score standardization, and then maps it to temporal feature vectors through a temporal encoder. S1054's multimodal feature projection, alignment, and concatenation inputs text vector T, image vector I, and temporal vector S into three independent linear projection layers, mapping them uniformly to the same semantic space. The aligned three vectors are concatenated sequentially to form a multidimensional fused multimodal feature vector F=[T',I',S'], where T', I', and S' represent each modality in the unified space. S1055, based on the associated enterprise identifier and stored in the feature library, binds the fused feature vector F to the unified social credit code of enterprise A. This record is written to the feature library, with fields including: enterprise ID, feature vector, generation time, industry, and associated risk dimension. This feature vector is subsequently called by S106 to predict the technology and innovation risk score and supports S107 in generating qualitative paragraphs such as: core technology concentrated on a single mRNA platform, recent Phase III clinical trial failure, indicating a risk of single technical route and translational failure. Ensure that each feature corresponds one-to-one with the enterprise to facilitate subsequent model calls and result traceability.

[0082] In one embodiment of the present invention, based on step S106, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S106 specifically includes the following: S1061: Performs standardized preprocessing on the fused multimodal feature vector output by S105, mapping the feature values ​​to the input numerical range of the generative AI risk prediction model, adapting to the input dimension and distribution requirements of the model.

[0083] The input values ​​can be in the range of 0-1 or a normal distribution with a mean of 0 and a variance of 1.

[0084] S1062: Call the generative AI risk prediction model that has been pre-optimized by domain adaptation and reinforcement learning. The generative AI risk prediction model takes standardized multimodal features as input and includes a feature encoding module, a risk dimension association module, and an output generation module.

[0085] Alternatively, one approach is to invoke a generative AI risk prediction model built on the Transformer architecture: The feature encoding module uses a multimodal encoder, such as the text-image encoding of CLIP and the temporal encoding of LSTM, to encode text, image and temporal features into a unified semantic vector.

[0086] The risk dimension association module is a learnable weight matrix, where each row corresponds to a risk dimension and each column corresponds to a feature dimension, storing the association strength between features and dimensions.

[0087] The output generation module includes a linear regression layer corresponding to the output score and a softmax classification layer corresponding to the output level.

[0088] The model is fine-tuned using domain data from S101 to adapt to the risk characteristics of the science and technology innovation field.

[0089] In this way, the multimodal encoder transforms heterogeneous text, image, and temporal features into a unified semantic space, eliminating modal differences; the risk dimension association module locates features related to the current risk dimension through a weight matrix; and the output module generates quantified scores and graded levels based on these features.

[0090] As one implementation method for training and optimizing the generative AI risk prediction model in step S1062, it specifically includes the following steps: S10621: Construct a training corpus for risk analysis of science and technology innovation enterprises, and load the domain risk feature configuration of S101 and the multi-source heterogeneous dataset of S104 as basic training samples.

[0091] S10622: Based on the general generative AI base model, the training corpus is used for domain-adaptive pre-training to update the model parameters to adapt to the semantic features of the risk domain of science and technology innovation enterprises.

[0092] S10623: Based on the pre-trained model, supervised fine-tuning is performed using a sample dataset labeled with scores for each risk dimension and the overall risk level to optimize the model's risk identification accuracy.

[0093] S10624: Configure a human feedback mechanism for risk analysts, and optimize model parameters through reinforcement learning strategies to make the generated risk scores and levels more in line with professional assessment standards.

[0094] S10625: Package the optimized model parameters, configuration information, and performance metrics into a deployable model asset and register it in the model library.

[0095] S1063: During model runtime, the risk dimension association module dynamically weights the encoded features using a predefined risk dimension-feature weight matrix.

[0096] For example, when dealing with technology and innovation risks, the weight of characteristics such as technological advancement and R&D investment intensity is strengthened, while irrelevant characteristics are suppressed.

[0097] In this embodiment, the risk dimension feature weight matrix is ​​obtained through domain-supervised learning. For example, for the risk dimension of clinical trial progress in the biomedical field, the weights of features such as clinical trial stage and number of enrolled patients in the matrix are set to 0.8, while the weight of positive public opinion rate is set to 0.1.

[0098] During model execution, the encoded feature vector is multiplied by the weight matrix to obtain a weighted feature vector. For example, in the weighted vector of technology and innovation risk, the value of the patent citation count feature is amplified, while the value of the executive turnover rate feature is reduced. This minimizes the interference of irrelevant features on risk assessment.

[0099] S1064: Based on the weighted feature vector, the output generation module calculates the continuous risk score for each risk dimension through a linear regression layer and outputs the risk level through a Softmax classification layer.

[0100] In this embodiment, the regression layer uses a linear activation function to output a continuous score from 0 to 10. A higher score indicates greater risk; for example, a score of 7.5 for R&D intensity indicates a high risk in this dimension. The classification layer uses a softmax function to output probability distributions for low, medium, and high risk categories, taking the category with the highest probability as the overall risk level. This provides a quantified risk score, facilitating horizontal comparisons of risk levels across different companies and dimensions.

[0101] S1065: Sort the output scores of each risk dimension and the overall level according to the risk dimension order defined in the reporting framework in S102 to generate risk scores and risk level results.

[0102] This embodiment follows the risk dimension order defined in the reporting framework of S102, such as executive summary, enterprise profile, industry and market risk, technology and innovation risk, etc., and sequentially fills the scores and levels output in S1064 into the corresponding positions to generate the result data. For example: {"Industry and Market Risk":{"Score":7.2,"Level":"Medium Risk"},"Technology and Innovation Risk":{"Score":8.1,"Level":"High Risk"},...}.

[0103] For each risk dimension, the contribution of key features in the model output is recorded simultaneously. For example, in the technology and innovation risk, the number of times patents are cited contributes 40% of the score. This serves as the source of evidence for the qualitative analysis in S107, ensuring that the risk results and evidence of each dimension can be accurately cited when S107 generates the qualitative analysis.

[0104] The following example, using a biomedical technology startup company A, and following steps S10621 to S10625, illustrates the process. S10621 involves constructing a risk analysis training corpus for startup companies, incorporating heterogeneous raw data from multiple biomedical companies from S104. Each company is associated with the subdivided domain labels output in S101. The corpus is organized as {Company ID, Domain Label, Raw Multimodal Data}, serving as the foundation for unsupervised / self-supervised pre-training. S10622 uses a general generative AI foundation model, such as DeepSeek-Coder or ChatGLM, as a basis. Domain pre-training is performed on a subset of biomedical data, with tasks including: Masking recovery clinical terminology (e.g., [MASK] phase clinical trial failure → Phase III); Comparative learning: narrowing the vector distance between CAR-T and T cell receptors and cytokine storms, while distancing it from concepts unrelated to small molecule chemical drugs; After pre-training, the model's understanding of professional contexts such as IND approval, BLA submission, and off-target toxicity was significantly enhanced, and it was able to accurately distinguish the risk level differences between clinical pauses and clinical terminations.

[0105] S10623 uses samples from multiple biopharmaceutical companies labeled by a pharmaceutical investment bank's risk control team; each sample includes: the standardized multimodal feature vector output by S105. Human scoring: e.g., Technology and Innovation Risk: 88 points, Product and Commercialization Risk: 92 points, Overall Level: High Risk. Fine-tuning objective: to make the model output score approximate the expert score. After fine-tuning, on the validation set, the MAE for commercialization risk prediction of companies with a single product dependency decreased to 3.8 points, and the Softmax F1-score for the high-risk category reached 0.93. S10624 generates a product and commercialization risk score of 75 points for company A; combined with the risk analyst's assessment, it is considered to be 90 points. The model is guided to focus on product pipeline breadth and clinical stage distribution characteristics; after optimization, the model learns to increase the commercialization risk to ≥85 points when Phase III failure occurs or there is no backup pipeline. S10625 outputs the optimized model parameters and input normalization configuration. In one embodiment of the present invention, based on step S107, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S107 specifically includes the following methods: S1071: Read the risk assessment result table output by S106 and the dimension attribute table containing risk dimension names and core analysis objectives in the modular report framework of S102, and generate an input list by associating them through risk dimension IDs.

[0106] S1072: Based on the risk dimension ID in the generated input list and the indicator ID in the indicator configuration package in S103, retrieve evidence data related to each dimension and indicator.

[0107] S1073: Load the text generation model file based on the retrieval-enhanced generation architecture, read the qualitative description specifications in the S102 report framework and the domain expression style in the S101 risk feature configuration, and write the specifications and style parameters into the model's generation constraint module through the model configuration tool.

[0108] It should be noted that the specific implementation method of writing specification and style parameters into the model's generation constraint module through the model configuration tool in S1073 is as follows: S10731: Load the pre-trained text generation model file, including the model structure definition file and weight parameter file, and initialize the model instance.

[0109] S10732: Read qualitative description specification data from the S102 report framework configuration, including industry terminology library files and paragraph format template files.

[0110] S10733: Read the domain expression style parameters from the S101 risk feature configuration, including a list of domain-specific terms and expression style rules.

[0111] S10734: The qualitative description specifications and domain representation style parameters are parsed by the model configuration tool and converted into a configuration format that can be recognized by the model generation constraint module.

[0112] S10735: Write the converted configuration parameters into the model's generation constraint module and update the generation parameter settings of the model instance.

[0113] It's important to note that reading qualitative description specifications and domain style parameters ensures that the model-generated text conforms to both general report format requirements and the professional terminology specific to the scientific and technological innovation field. Parameter transformation and writing ensure that external rules are effectively received and applied by the model, maintaining consistency and professionalism in the format, terminology, and style of the generated qualitative analysis paragraphs, thereby enhancing the report's credibility and readability.

[0114] S1074: Combine the dimension names, risk scores, and evidence data from the corresponding evidence folder in the generated input list into the model input text according to the preset format. The configured text generation model will generate qualitative analysis paragraphs for each risk dimension.

[0115] Optionally, the dimension names and risk scores from the generated input list are concatenated with the evidence data from the corresponding evidence folder in a fixed format, such as [Dimension Name] Risk Assessment: The overall score is [Score] points. The main evidence includes: 1. [Content of Evidence 1]; 2. [Content of Evidence 2]..., forming input text that the model can recognize. The input text is then passed to the configured text generation model. The model outputs qualitative analysis paragraphs based on the retrieved evidence and generation constraints. These paragraphs are stored as TXT files, with the filename containing the risk dimension ID.

[0116] S1075: Based on the S102 modular reporting framework and S103 indicators, perform format validation on the generated qualitative analysis paragraphs. Paragraphs that pass the validation are stored as a paragraph set according to the risk dimension ID.

[0117] This embodiment performs a step-by-step check on the generated paragraph files based on the description standards of the S102 reporting framework and S103 indicators. If any formatting is missing, it is marked as a formatting error. If there are logical contradictions, it is marked as a logical error. Paragraphs that pass the check are sorted by risk dimension ID to form a paragraph set. This ensures that the generated content is consistent with the risk score, evidence, and framework logic.

[0118] It should be noted that the optimization training process of the text generation model can extract historical report texts and corresponding risk data from the enterprise knowledge base to construct the original training corpus and configure a standardized training sample set. The model is then trained in batches using this training sample set. The difference between the output and sample labels is calculated through forward propagation, and the model weight parameters are updated using the backpropagation algorithm. A learning rate scheduler is used to dynamically adjust the parameter update step size during training. The model is fine-tuned using a specialized dataset with format annotations, and a constraint module forces the model output to conform to paragraph templates and terminology requirements. Gradient pruning is used during training to prevent parameter overflow and intermediate checkpoints are retained for fault recovery. A quality score is generated as a reward signal by comparing the generated content with the standard template. Based on the reward signal, the model parameters are adjusted using a policy gradient method to iteratively optimize the generation quality. The optimization process is accelerated using a distributed computing architecture. The performance of the optimized model is evaluated using a validation dataset, and automated test scripts are used to check the output format compliance and content relevance. The final model weights, constraint module configuration, and version information are packaged into a model file package, registered in the model repository, and a digital signature is generated to ensure integrity.

[0119] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0120] like Figure 2 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of a risk report analysis method for science and technology innovation enterprises based on generative AI.

[0121] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0122] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0123] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0124] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0125] The communication module 104 transmits radio signals to and / or receives radio signals from at least one of a base station, an external terminal, and a server. Such radio signals may include voice call signals, video call signals, or various types of data sent and / or received according to text and / or multimedia messages.

[0126] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the generative AI-based risk report analysis method for science and technology innovation enterprises.

[0127] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0128] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the exemplary methods section of this specification according to various exemplary embodiments of this disclosure.

[0129] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for analyzing risk reports of science and technology enterprises based on generative AI, characterized in that, The method comprises: S101: identifying the scientific and innovative field to which the enterprise belongs, and obtaining the risk feature configuration of the scientific and innovative field; S102: defining a modular report framework, including multiple risk dimensions, and configuring a data requirement list and feature extraction rules for each dimension; S103: for each risk dimension in the report framework, define quantitative indicators and qualitative indicators based on the risk feature configuration obtained in S101; S104: based on the data requirement list of S102 and the indicators defined in S103, obtain enterprise data from a database, and construct a multi-source heterogeneous data set through ETL; S105: based on the feature extraction rules of S102 and the indicators defined in S103, extract text, image and time series data features from the multi-source heterogeneous data set, and perform feature fusion to obtain fused multi-modal features; S106: input the fused multi-modal features obtained in S105 into a generative AI model for enterprise risk prediction tasks, and output risk scores and overall risk levels for each risk dimension; S107: based on the risk scores and risk levels obtained in S106, combined with relevant evidence retrieved from a knowledge base, use a text generation model based on retrieval-enhanced generation architecture to generate qualitative analysis paragraphs, and assemble them into a complete risk report according to the report framework defined in S102; S108: display the complete risk report generated in S107 in a visual manner.

2. The sci-tech enterprise risk report analysis method based on generative AI according to claim 1, characterized in that, S102 specifically includes the following methods: Establish a tree structure risk dimension framework containing multiple first-level risk dimensions and their subordinate second-level sub-dimensions; For each second-level sub-dimension in the risk dimension framework, define a standardized data requirement list containing data types, data sources and data fields; Based on the data requirement list, configure text feature extraction methods, image feature extraction methods and time series data feature extraction methods for each second-level sub-dimension; Construct a data association relationship mapping table and a feature interaction relationship mapping table between the second-level sub-dimensions in the risk dimension framework.

3. The generative AI-based sci-tech enterprise risk report analysis method according to claim 1, characterized in that, S103 specifically includes the following methods: Call the risk feature configuration output by S101, load the core indicator item set corresponding to the current scientific and innovative field; Based on the risk dimensions in the report framework defined in S102, select and bind at least one quantitative indicator and at least one qualitative indicator for each dimension from the core indicator item set; For each bound quantitative indicator, configure the required data source field and calculation formula expression; For each bound qualitative indicator, configure the data source and key identification required for analysis; Integrate the mapping relationship of all indicators, data sources and rules to generate an indicator configuration file.

4. The sci-tech enterprise risk report analysis method based on generative AI according to claim 1, characterized in that, S105 specifically includes the following methods: According to the feature extraction rules of S102 and the indicators defined in S103, dispatch different feature extractors from the multi-source heterogeneous data set to extract features from text, images and time series data respectively; Perform entity analysis and vectorization representation on the extracted text features to generate text feature vectors; Standardize and serialize the extracted image features and time series data features to generate image feature vectors and time series feature vectors; Projecting and aligning the text feature vector, the image feature vector and the time sequence feature vector based on a preset unified dimension space, and splicing to form a fused multi-modal feature vector; Correlate the fused multi-modal feature vector with the enterprise identifier.

5. The sci-tech enterprise risk report analysis method based on generative AI according to claim 1, characterized in that, S106 specifically includes the following modes: Standardizing the fused multi-modal feature vector output by S105, mapping the feature values to the input numerical range of the generative AI risk prediction model, and adapting the input dimension and distribution requirements of the model; Call the generative AI risk prediction model pre-adapted and optimized by domain and reinforcement learning, and the generative AI risk prediction model takes the standardized multi-modal features as input; During model running, the risk dimension correlation module dynamically weights the encoded features through a pre-defined risk dimension-feature weight matrix; Based on the weighted feature vector, the output generation module calculates the continuous risk scores of each risk dimension through a linear regression layer, and outputs the risk levels through a Softmax classification layer; Sort the output risk dimension scores and overall levels according to the risk dimension order defined in S102, and generate risk score and risk level results.

6. The sci-tech enterprise risk report analysis method based on generative AI according to claim 5, characterized in that, S1062 specifically includes the following modes: Construct a scientific and innovative enterprise risk analysis training corpus, load domain risk feature configuration and multi-source heterogeneous data set as basic training samples; Based on the general generative AI base model, use the training corpus for domain adaptability pre-training, update the model parameters to adapt to the semantic features of the scientific and innovative enterprise risk field; Based on the pre-trained model, use the sample data set labeled with risk dimension scores and overall risk levels for supervised fine-tuning to optimize the risk recognition accuracy of the model.

7. The sci-tech enterprise risk report analysis method based on generative AI according to claim 1, characterized in that, S107 specifically includes the following modes: Read the risk assessment result table output by S106 and the dimension attribute table containing risk dimension names and core analysis targets in the S102 modular report framework, and generate an input list through risk dimension ID correlation; Based on the risk dimension ID in the generated input list and the index ID in the index configuration package, retrieve evidence data related to each dimension and index; Load the text generation model file based on the retrieval-enhanced generation architecture, read the qualitative description specification in the S102 report framework and the domain expression style in the S101 risk feature configuration, and write the specification and style parameters into the model's generation constraint module through the model configuration tool; Combine the dimension names, risk scores in the generated input list with the evidence data in the corresponding evidence folder into model input text according to the preset format, and generate qualitative analysis paragraphs for each risk dimension using the configured text generation model.

8. The sci-tech enterprise risk report analysis method based on generative AI according to claim 7, characterized in that, S1073 specifically includes the following modes: Load the pre-trained text generation model file; Read the qualitative description specification data from the report framework configuration; Read the domain expression style parameters from the risk feature configuration, including the domain-specific term list and expression style rules; Parse the qualitative description specification and domain expression style parameters through the model configuration tool, and convert them into a configuration format recognizable by the model's generation constraint module; Write the converted configuration parameters into the generation constraint module of the model, and update the generation parameter settings of the model instance.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the scientific and creative enterprise risk report analysis method based on generative AI according to any one of claims 1-8 when executing the program.

10. A storage medium having stored thereon a computer program, characterized in that The computer program implements the steps of the scientific and creative enterprise risk report analysis method based on generative AI according to any one of claims 1-8 when executed by the processor.

Images