Legal risk early warning and compliance management methods and systems

Abstract

The disclosed embodiment provides a legal risk warning and compliance management method and system, wherein the legal risk warning and compliance management method includes: obtaining initial data related to laws and contracts, and preprocessing the initial data to generate structured analysis data; constructing a ternary data structure of the analysis data, and performing semantic analysis on the ternary data structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data; outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result; performing visual analysis processing on the initial data to generate a second initial risk assessment result; generating a risk assessment report based on the first initial risk assessment result and the second initial risk assessment result. The above technical solution can improve the efficiency and accuracy of legal risk warning and compliance management.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of data processing technology, and in particular to a legal risk warning and compliance management method and system. Background Art

[0002] With the continuous development of the rule of law society, the complexity of the business environment and the continuous updating of laws and regulations have led to increasing legal risks and contract compliance challenges faced by enterprises.

[0003] At present, traditional legal risk and contract compliance management methods mainly rely on manual review and fixed rule bases, which have significant defects. For example, human factors may lead to omissions or errors in judgment, especially when dealing with complex or ambiguous legal issues. The accuracy is limited, and manual review of large amounts of documents and data is time-consuming and labor-intensive, making it difficult to meet the needs of real-time monitoring.

[0004] In this context, how to provide technical solutions to improve the efficiency and accuracy of legal risk warning and compliance management has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the embodiments of the present disclosure provide a legal risk warning and compliance management method and system, which can improve the efficiency and accuracy of legal risk warning and compliance management.

[0006] The present disclosure provides a legal risk warning and compliance management method, including:

[0007] Acquire initial data related to the law and the contract, and pre-process the initial data to generate structured analysis data, wherein the initial data includes at least one of text data, image data, and audio data;

[0008] Constructing a ternary metadata structure of the analysis data, and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data;

[0009] Outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result;

[0010] Performing visual analysis processing on the initial data to generate a second initial risk assessment result, including: using a pre-trained YOLOv5 model to identify structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements; using a ResNet model to extract the initial data for semantic analysis, and extracting key information in the image area corresponding to the initial data; using an attention mechanism to fuse the logical structure tree and the key information, and based on the confidence scores of the logical structure tree and the key information, adjusting the fusion ratio of the logical structure tree and the key information to obtain visual information corresponding to the initial data; converting the visual information to obtain a sequence representation of the visual information; performing a structured analysis on the sequence representation of the visual information to identify the risk points and confidence levels of the sequence list; and generating a second initial risk assessment result based on the risk points and confidence levels of the sequence list;

[0011] A risk assessment report is generated according to the first initial risk assessment result and the second initial risk assessment result.

[0012] Optionally, the constructing of the ternary metadata structure of the analysis data, and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data, includes:

[0013] Generating a ternary systolic array includes: defining a circular buffer having a fixed size, and defining ternary data in the circular buffer;

[0014] Adopting a cyclic writing strategy and a time window mechanism, the analysis data acquired within a specific time is input into the ternary systolic array, and a ternary data structure corresponding to the type of the analysis data is generated, wherein the ternary data structure includes: data, timestamp and credibility;

[0015] The ternary data structures are input into a pre-trained multimodal language model in batches, and semantic analysis is performed on the ternary data structures of the analysis data through the multimodal language model to identify key elements in the ternary data structures of the analysis data, and analyze the key elements to determine data with potential risk points in the analysis data and give corresponding risk scores.

[0016] Optionally, the step of adopting a cyclic write strategy and a time window mechanism to input the acquired analysis data within a specific time into the ternary systolic array further includes:

[0017] A weighted average or aggregation algorithm is used to determine the credibility of the analysis data, and analysis data with a credibility lower than a set credibility is filtered out.

[0018] Optionally, the multimodal language model is obtained in the following manner:

[0019] Selecting a multi-objective transformer model, integrating a multimodal processing module on the multi-objective transformer model, and configuring an attention mechanism on the multi-objective transformer model;

[0020] Collect legal text corpus and contract text forecasts, as well as pre-constructed legal risk identification statements and compliance assessment data sets, train the multi-objective transformer model, and generate the multimodal language model.

[0021] Optionally, in the case where the initial data is audio data and / or video data, after converting the audio data and / or video data into corresponding text data and obtaining converted text data, the ternary systolic array is used to construct a ternary data structure of the converted text data;

[0022] Wherein, when the initial data is audio data, the ternary data structure includes: a text segment, a timestamp, and an ASR confidence level;

[0023] When the initial data is image data, the ternary data structure includes: a text segment, a timestamp, and an OCR confidence level.

[0024] Optionally, the Bayesian network includes: an input node, a hidden node and an output node, wherein the hidden node is provided with a conditional probability table generated based on the empirical coding of the target population and an evaluation data set for characterizing various legal scenarios and different types of contracts;

[0025] Outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result includes:

[0026] The target data is discretized through the input node to obtain discrete data, and output to the hidden node;

[0027] By using the hidden nodes, based on the discrete data, the conditional probability table and the evaluation data set, an expectation maximization algorithm is applied to generate the first initial risk assessment result;

[0028] And outputting the first initial risk assessment result through the output node.

[0029] Optionally, the step of discretizing the target data through the input node to obtain discrete data further includes:

[0030] A noise gate mechanism is set on the input node, and target data having a risk score lower than a preset risk score corresponding to the noise gate mechanism is filtered out.

[0031] Optionally, performing visual analysis on the initial data to generate a second initial risk assessment result includes:

[0032] Using a pre-trained YOLOv5 model, identifying structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements;

[0033] The ResNet model is used to extract the initial data for semantic analysis, and key information in the image area corresponding to the initial data is extracted;

[0034] Applying an attention mechanism to fuse the logical structure tree and the key information, and adjusting the fusion ratio of the logical structure tree and the key information based on the confidence scores of the logical structure tree and the key information to obtain visual information corresponding to the initial data;

[0035] Converting the visual information to obtain a sequence representation of the visual information;

[0036] Performing structured analysis on the sequence representation of the visual information to identify risk points and confidence levels of the sequence table;

[0037] Based on the risk points and confidence levels in the sequence table, a second initial risk assessment result is generated.

[0038] Optionally, generating a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result includes:

[0039] Use text-image alignment technology to establish the spatial correspondence between text risk points and visual risk elements;

[0040] Using multimodal fusion technology, constructing a joint representation including text and visual features for each risk point in the first initial risk assessment result and the second initial risk assessment result;

[0041] Use graph convolutional networks or Transformers to capture long-distance dependencies between risk points;

[0042] Using a comprehensive scoring model, based on the spatial correspondence, the joint representation, and the long-distance dependency, to generate a risk score and confidence level for each risk point;

[0043] Use hierarchical clustering or topic modeling techniques to group risk points and generate a document-level risk summary, which includes a risk category, an overall risk score, and a list of key risk points;

[0044] Represent the dependencies between different risk factors based on probabilistic graphical models;

[0045] Implement a feedback loop mechanism to dynamically adjust the weight of text and visual analysis based on historical accuracy and generate natural language explanations for each integrated risk assessment result;

[0046] Using visualization technology, the risk assessment report is generated.

[0047] The present disclosure also provides a legal risk warning and compliance management system, including:

[0048] A data acquisition module, adapted to acquire initial data related to laws and contracts, and pre-process the initial data to generate structured analysis data, wherein the initial data includes at least one of text data, image data, and audio data;

[0049] A semantic analysis module, adapted to construct a ternary metadata structure of the analysis data, and perform semantic analysis on the ternary metadata structure of the initial data, identify data with potential risk points in the analysis data, and give corresponding risk scores as target data;

[0050] An inference module, adapted to output the target data to a pre-trained Bayesian network, perform inference processing on the target data, and generate a first initial risk assessment result; and adapted to perform visual analysis processing on the initial data to generate a second initial risk assessment result, comprising: using a pre-trained YOLOv5 model to identify structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements; using a ResNet model to extract the initial data for semantic analysis, and extracting key information in the image area corresponding to the initial data; using an attention mechanism to fuse the logical structure tree and the key information, and based on the confidence scores of the logical structure tree and the key information, adjusting the fusion ratio of the logical structure tree and the key information to obtain visual information corresponding to the initial data; converting the visual information to obtain a sequence representation of the visual information; performing a structured analysis on the sequence representation of the visual information to identify the risk points and confidence levels of the sequence table; and generating a second initial risk assessment result based on the risk points and confidence levels of the sequence table;

[0051] A generating module is adapted to generate a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result.

[0052] Compared with the prior art, the technical solution of the embodiment of the present disclosure has the following advantages:

[0053] By adopting the legal risk warning and compliance management method provided by the embodiment of the present disclosure, by preprocessing the acquired initial data related to the law and the contract, structured analysis data can be generated, the consistency of the data can be improved, and the difficulty of generating the ternary data structure of the analysis data can be reduced, and the ternary data structure can better reflect the legal significance contained in the analysis data, so that by semantic analysis of the ternary data structure, the data with potential risk points in the analysis data can be accurately identified, and by giving corresponding risk scores to the data with potential risk points, data can be selected from them as target data, so that the amount of data calculation can be reduced, and then based on the pre-trained Bayesian network, the accuracy of the risk score can be further improved, and then the accuracy of the generated first initial risk assessment result can be improved; at the same time, by visual analysis and processing of the initial data, the comprehensibility of the initial data can be enhanced, and then the accuracy of the generated second initial risk assessment result can be improved, so that when a risk assessment report is generated based on the first initial risk assessment result and the second initial risk assessment result, the efficiency and accuracy of legal risk warning and compliance management can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments of the present disclosure or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A flowchart of a legal risk early warning and compliance management method in an embodiment of the present disclosure is shown;

[0056] Figure 2 A flow chart of a target data selection process in an embodiment of the present disclosure is shown;

[0057] Figure 3 A flowchart of a process for generating a first initial risk assessment result in an embodiment of the present disclosure is shown;

[0058] Figure 4 A structural diagram of a legal risk warning and compliance management system in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0059] As mentioned in the background technology, traditional legal risk and contract compliance management methods mainly rely on manual review and fixed rule bases, which have significant defects.

[0060] As language model technology has made breakthrough progress in the field of natural language processing, it has provided new possibilities for solving the above problems. However, directly applying existing language model technology to legal risk warning and compliance management still faces many challenges, such as: 1) the legal field is highly professional, and general language models may lack sufficient domain knowledge; 2) legal texts usually contain complex logical relationships and contextual dependencies, requiring stronger reasoning capabilities; 3) legal risk assessment needs to consider multimodal information (such as text, images, audio), rather than just text analysis; 4) the "hallucination" problem of language models may lead to serious consequences in the legal field and contracts, and requires special treatment; 5) legal risk warning requires extremely high accuracy and interpretability, which places higher demands on the output quality of the model.

[0061] In order to solve the above technical problems, the embodiment of the present disclosure provides a legal risk warning and compliance management method. By preprocessing the acquired initial data related to the law and the contract, structured analysis data can be generated to improve the consistency of the data, thereby reducing the difficulty of generating a ternary data structure of the analysis data, and the ternary data structure can better reflect the legal significance contained in the analysis data. In this way, by performing semantic analysis on the ternary data structure, data with potential risk points in the analysis data can be accurately identified, and by giving corresponding risk scores to the data with potential risk points, data can be selected from them as target data, thereby reducing the amount of data calculation, and then based on the pre-trained Bayesian network, the accuracy of the risk score can be further improved, thereby improving the accuracy of the generated first initial risk assessment result; at the same time, by performing visual analysis processing on the initial data, the comprehensibility of the initial data can be enhanced, thereby improving the accuracy of the generated second initial risk assessment result, so that when a risk assessment report is generated based on the first initial risk assessment result and the second initial risk assessment result, the efficiency and accuracy of legal risk warning and compliance management can be improved.

[0062] That is, to achieve innovative solutions for efficient, accurate and real-time legal risk warning and compliance management.

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is described below with reference to the accompanying drawings.

[0064] See also Figure 1 A flowchart of a legal risk warning and compliance management method in an embodiment of the present disclosure is shown in FIG. Figure 1 As shown, the following steps can be performed:

[0065] S11, obtaining initial data related to laws and contracts, and preprocessing the initial data to generate structured analysis data.

[0066] Specifically, legal documents and contract documents are diverse and complex, so a large amount of initial data related to law and contracts can be obtained.

[0067] More specifically, multiple types of different initial data related to law and contracts can be obtained, that is, multimodal information can be obtained to improve the accuracy of subsequent risk assessment reports.

[0068] In some embodiments, the initial data may include at least one of text data (such as contracts, emails, internal documents, etc.), image data (such as scans of signed documents, on-site photos), and audio data (such as meeting minutes, customer communication recordings).

[0069] It is to be understood that the types of initial data described in the above examples are merely illustrative and are used to indicate the acquisition of different types of data, and should not be construed as limiting the present invention. For example, in some other embodiments, the initial data may also include video data.

[0070] In some embodiments of the present disclosure, considering the differences between different initial data, the initial data may be preprocessed to generate structured analysis data, which can unify the format of the initial data and reduce the difficulty of subsequent processing.

[0071] In some examples, the initial data can be cleaned and standardized, for example, to remove noise and redundant information, unify the format, and desensitize the data to protect privacy information.

[0072] Furthermore, different types of initial data are processed into structures, for example, structured extraction of text data, OCR (Optical Character Recognition) processing of image data, and transcription and structuring of structured and audio data.

[0073] S12, constructing a ternary metadata structure of the analysis data, and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data.

[0074] Specifically, after step S11, structured analysis data can be generated, and by constructing a ternary data structure for these structured analysis data, the legal significance contained in the analysis data can be more truly reflected. In this way, when performing semantic analysis on the ternary data structure of the initial data, data with potential risk points in the analysis data can be more accurately identified, and by giving corresponding risk scores to the data with potential risk points, data can be selected from them as target data, thereby reducing the amount of data calculation.

[0075] That is, the target data includes data with potential risk points and their risk scores.

[0076] S13, outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result.

[0077] Specifically, the target data represents data with potential risk points. By outputting the target data to a pre-trained Bayesian network, and the Bayesian network has better generalization and robustness, the target data can be inferred and processed more realistically and accurately. The first initial risk assessment result generated can more realistically reflect whether the data with potential risk points is real risk data.

[0078] S14, performing visual analysis processing on the initial data to generate a second initial risk assessment result.

[0079] Specifically, the visual analysis processing can enhance the comprehensibility of the initial data, so that when performing risk assessment on the initial data, the accuracy of the generated second initial risk assessment result can be improved.

[0080] S15: Generate a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result.

[0081] Specifically, the first initial risk assessment result is obtained through semantic analysis and model prediction of the analysis data, while the second initial risk assessment result is obtained based on visual analysis processing. That is, this solution adopts two completely different analysis methods to analyze the same data, thereby effectively avoiding the limitations of a single analysis method and better capturing the risk data in the initial data, thereby improving the efficiency and accuracy of legal risk warning and compliance management.

[0082] It should be noted that when executing the above steps, the order of the steps can be flexibly changed. For example, as long as steps S11 to S14 are executed before step S15.

[0083] In short, the embodiments of the present disclosure can improve the efficiency and accuracy of legal risk warning and compliance management by performing preprocessing, ternary data structure, semantic analysis and prediction processing, and visual analysis processing on legal documents.

[0084] In some embodiments of the present disclosure, after the initial data is preprocessed to generate structured analysis data, semantic analysis processing may be performed on the analysis data to identify data with potential risk points.

[0085] As an example, see Figure 2 A flowchart of a target data selection process in an embodiment of the present disclosure is shown in FIG. Figure 2 As shown, the following steps can be performed:

[0086] S21, generate a ternary systolic array.

[0087] The step of generating a ternary systolic array includes: defining a circular buffer with a fixed size (i.e., a systolic array architecture to achieve efficient parallel processing), and defining ternary data in the circular buffer (i.e., a data structure form of the analyzed data after being processed by the circular buffer).

[0088] S22, adopting a cyclic write strategy and a time window mechanism, inputting the analysis data acquired within a specific time into the ternary systolic array, and generating a ternary data structure corresponding to the type of the analysis data.

[0089] Specifically, by adopting a circular write strategy, a real-time update mechanism for data streams can be implemented. This is because: the circular write strategy can ensure that the latest data is always available and can implement atomic operations to ensure data consistency under high concurrency; and by adopting a time window mechanism, a timestamp-based data synchronization strategy can be implemented. This is because: using the time window mechanism, only analytical data within a specific time range is processed, and orderly processing of analytical data is achieved, ensuring that causal relationships are not destroyed, thereby improving the timeliness of analytical data.

[0090] In this way, the analysis data processed by the ternary systolic array can obtain a ternary data structure that reflects the analysis data in real time.

[0091] In some embodiments, the triple data structure may include data, timestamp, and credibility, wherein the data may represent the actual risk score or other relevant information, the timestamp may represent the time when the recorded data is generated or updated, and the credibility may represent the reliability or importance of the data.

[0092] In some embodiments of the present disclosure, the specific expression and generation of the ternary data structure may be different depending on the type of analysis data.

[0093] For example, when the initial data is audio data and / or video data, after the audio data and / or video data is converted into corresponding text data and the converted text data is obtained, a ternary systolic array is used to construct a ternary data structure of the converted text data.

[0094] Among them, when the initial data is audio data, the three-dimensional data structure includes: text segment, timestamp, ASR confidence; when the initial data is image data, the three-dimensional data structure includes: text segment, timestamp, OCR confidence.

[0095] S23, inputting the ternary data structure into a pre-trained multimodal language model in batches, performing semantic analysis on the ternary data structure of the analysis data through the multimodal language model, identifying key elements in the ternary data structure of the analysis data, and analyzing the key elements, determining data with potential risk points in the analysis data, and giving corresponding risk scores.

[0096] In some embodiments of the present disclosure, a multimodal language model is obtained in the following manner: a multi-objective transformer model is selected, a multimodal processing module is integrated into the multi-objective transformer model, and an attention mechanism is configured on the multi-objective transformer model; legal text corpus and contract text expectations, as well as pre-constructed legal risk identification statements and compliance assessment data sets are collected, the multi-objective transformer model is trained, and the multi-objective transformer model is generated.

[0097] More specifically, the model architecture design includes: designing a cross-task multi-objective transformer architecture, integrating a multimodal input processing module, and designing a legal field-specific attention mechanism.

[0098] Model pre-training, including: collecting large-scale legal text corpus, designing pre-training tasks, including masked language model and next sentence prediction, and implementing large-scale distributed pre-training.

[0099] Finally, fine-tuning in the legal field includes: building legal risk identification and compliance assessment datasets, designing multi-task fine-tuning strategies, and performing model fine-tuning and verification.

[0100] After the above steps, a multimodal language model that meets the requirements of the legal and contractual fields can be obtained, so that when performing semantic analysis on the ternary data structure, the key elements in the ternary data structure of the analysis data (for example, data with sensitive information and abnormal expressions) can be more accurately identified. Then, by analyzing the key elements, a focusing effect can be achieved, so that data with potential risk points in the analysis data can be determined and a corresponding risk score (that is, the probability that the data is risky data) can be given.

[0101] Furthermore, by inputting the ternary data structures into the multimodal language model in batches, a balance between real-time performance and computational efficiency can be achieved.

[0102] In some optional embodiments of the present disclosure, a cyclic write strategy and a time window mechanism are used to input the analysis data acquired within a specific time into the step of the ternary pulsation array, which may also include: using a weighted average or aggregation algorithm to determine the credibility of the analysis data, and filtering out the analysis data with a credibility lower than the set credibility, so that interference data can be filtered in advance to improve the accuracy of risk scoring.

[0103] Therefore, by integrating the ternary pulsation array with the multimodal language model, it is possible to comprehensively consider complex scenarios with multiple factors and constantly updated information, thereby significantly improving the response speed and analysis accuracy, and providing enterprises with timely and reliable legal risk warnings.

[0104] For ease of understanding, an example is used to illustrate the data stream processing process based on the ternary systolic array.

[0105] A1) Designing a ternary systolic array SA, including: defining a ternary data structure and a systolic array architecture, wherein the systolic array architecture creates a fixed-size circular buffer to achieve efficient parallel insert and read operations, and designs a data format conversion function to convert the systolic array output into a format that can be processed by a multimodal language model, and implements a batch processing mechanism to balance real-time performance and computational efficiency.

[0106] A2) Data stream algorithm implementation, including: using a circular write strategy to implement a real-time update mechanism for data streams; using a time window mechanism to implement a timestamp-based data synchronization strategy to process analytical data within a specific time range; and optionally, a weighted average or aggregation method to determine the credibility of analytical data and set a credibility threshold to filter out low-quality analytical data.

[0107] A3) Integrate the ternary systolic array SA with the multimodal language model, including: using a data structure conversion function to convert the ternary systolic array SA output into a format that can be processed by the multimodal language model, and ensuring that key information (such as time order and credibility) is retained during the conversion process; dynamically adjust the batch size according to the data flow speed and the processing capacity of the multimodal language model, and use a priority queue to ensure that high-confidence or time-sensitive analysis data is processed first.

[0108] and optionally, A4) feeding the output of the multimodal language model into the ternary systolic array SA to balance new and old information.

[0109] As an example, suppose we are developing a real-time contract review system that needs to process continuous data streams from multiple sources, including contract text, market data, and regulatory updates. We will use a ternary systolic array SA to manage these data streams and provide high-quality input for a multimodal language model.

[0110] Data flow settings: Contract text flow: There may be multiple paragraphs that need to be analyzed per second; Market data flow: Update relevant market indicators every minute; Regulatory update flow: Receive the latest changes in laws and regulations from time to time.

[0111] The ternary pulsation array SA constructs a ternary data structure, in which the contract text array: [text fragment, timestamp, confidence of the NLP model]; the market data array: [indicator value, timestamp, data source reliability]; the regulatory update array: [regulatory content, effective time, degree of impact].

[0112] Data processing flow, including: when new contract text paragraphs arrive, insert them into the contract text array; regularly update the market data array; and immediately insert the regulatory update array;

[0113] Data synchronization and filtering: Use time windows (such as the last 5 minutes) to select relevant data, and filter low-quality data based on confidence thresholds.

[0114] Finally, aggregation operations are performed, including: contract text: merge relevant paragraphs and generate a summary; market data: calculate weighted average; regulatory updates: prioritize high-impact updates.

[0115] The aggregated data is converted into an input format acceptable to the multimodal language model, and the data batch size is adjusted according to the processing capacity of the multimodal language model.

[0116] Then, the multimodal language model receives the processed data, performs real-time risk assessment, and feeds the assessment results back to the ternary systolic array SA to update the credibility of the relevant data.

[0117] Finally, the output example is: "High-risk clauses detected: Article 7 on limitation of liability may violate the latest amendment to the Consumer Rights Protection Law (credibility 0.92). It is recommended to revise the clauses to comply with the new regulations, otherwise you may face a fine of up to 5 million yuan (based on the latest market data, this amount is approximately 5% of the company's annual revenue)".

[0118] This example shows how the ternary systolic array SA can effectively manage and process real-time data streams from multiple sources. It can: ensure the timeliness of data, ensure the processing of the latest information through timestamps; improve data quality, filter unreliable information through credibility mechanisms; achieve efficient data aggregation, and provide high-quality input for multimodal language models; support real-time feedback, allowing the output of multimodal language models to dynamically update information in the data stream.

[0119] Using the above example, data including potential risk points with risk scores can be generated based on the ternary systolic array SA and the multimodal language model. At this time, a first initial risk assessment result for characterizing the true risk level of the target data can be generated through the Bayesian network.

[0120] In some embodiments of the present disclosure, a Bayesian network may include: input nodes, hidden nodes, and output nodes.

[0121] Among them, the input node is used to receive the output result of the multimodal language model, the hidden node is used to represent the intermediate reasoning process and reasoning result, and the output node represents the first initial risk assessment result.

[0122] In some embodiments, the hidden nodes are provided with a conditional probability table generated based on the empirical coding of the target population and an evaluation data set for characterizing various legal scenarios and different types of contracts.

[0123] In this case, see Figure 3 A flowchart of a process for generating a first initial risk assessment result in an embodiment of the present disclosure is shown in FIG. Figure 3 As shown, the following steps can be performed:

[0124] S31, discretizing the target data through the input node to obtain discrete data, and outputting the discrete data to the hidden node.

[0125] S32, applying an expectation-maximization algorithm through the hidden nodes based on the discrete data, the conditional probability table and the evaluation data set to generate the first initial risk assessment result.

[0126] S33: Output the first initial risk assessment result through the output node.

[0127] Therefore, through the above steps, through the combination of the Yes network and the multimodal language model, as well as the conditional probability table generated by the experience encoding of the target population and the evaluation data set of various legal scenarios and different types of contracts, it is possible to provide a more accurate and reliable risk assessment, and refine and correct the risk assessment based on the experience of the target population, which is particularly suitable for processing legal texts, which require a high degree of expertise and careful judgment.

[0128] In some optional embodiments, the step of discretizing the target data through the input node to obtain discrete data may also include: setting a noise gate mechanism on the input node, and filtering out target data with a risk score lower than a preset risk score corresponding to the noise gate mechanism, so as to further improve the accuracy of risk identification.

[0129] In some examples, a threshold function is set to filter target data output by the modal language model that is lower than a set confidence level, where the threshold function is:

[0130] f ( x )= x , if x > threshold , else 0.

[0131] To facilitate understanding, an example is used to illustrate the combination process between the Yes network and the multimodal language model.

[0132] Suppose we are analyzing a complex business contract. The multimodal language model has performed a preliminary analysis of the contract. Now we want to use the noise gate Bayesian network to further improve the accuracy of risk identification.

[0133] First, the multimodal language model identified a potentially risky term and assigned a risk score of 0.7 (where 1 represents the highest risk and 0 represents no risk).

[0134] Second, if the noise gate threshold is set to 0.6, since the output of the multimodal language model (0.7) is higher than the threshold, this output is retained, that is, the noise gate output = 0.7

[0135] Third, the Bayesian network inference, that is, receiving the output of the noise gate (0.7), is discretized into a high-risk input (1), and then based on the input and prior knowledge, the intermediate inference result is calculated to finally generate the first initial risk assessment result, where the inference result of the first initial risk assessment result is a risk probability of 0.8.

[0136] Fourth, explanation of the results: The output of the multimodal language model indicated a medium to high risk (0.7), and after being processed by the noise-gate Bayesian network, the risk assessment increased to 0.8, which suggests that the Bayesian network, based on prior knowledge and expert experience, believes that this clause may have a higher risk than the initial assessment of the multimodal language model.

[0137] Fifth, a detailed risk report was generated, stating: "A high-risk clause was found in Clause Y, Section X, with an initial risk score of 0.7 and a risk probability of 0.8 after expert system analysis. It is recommended that the legal team carefully review the clause, especially the part on limitation of liability."

[0138] In some embodiments of the present disclosure, a semantic analysis and model prediction method can be used to generate a first initial risk assessment result. To improve the accuracy of the subsequent risk assessment report, the initial data can also be visually analyzed to generate a second initial risk assessment result.

[0139] Specifically, the method includes: using a pre-trained YOLOv5 model to identify structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements; using a ResNet model to extract the initial data for semantic analysis, and extracting key information in the image area corresponding to the initial data; using an attention mechanism to fuse the logical structure tree and the key information, and adjusting the fusion ratio of the logical structure tree and the key information based on the confidence scores of the logical structure tree and the key information to obtain visual information corresponding to the initial data; converting the visual information to obtain a sequence representation of the visual information; performing a structured analysis on the sequence representation of the visual information to identify the risk points and confidence levels of the sequence list; and generating a second initial risk assessment result based on the risk points and confidence levels of the sequence list.

[0140] More specifically, first, multi-model collaborative visual information extraction, including: dedicated visual model selection and training and model collaboration mechanism design, among which, dedicated visual model selection and training include: selecting OCR models suitable for legal documents and contract texts, training document layout analysis models, and image content understanding models; model collaboration mechanism design includes: designing model output fusion strategies, confidence-based decision-making mechanisms, and conflict resolution algorithms.

[0141] And optionally, the visual information of the multi-model collaboration is interacted with the multimodal language model to convert the visual information into a representation understandable by the multimodal language model, thereby achieving alignment of the visual information with the textual information.

[0142] In some embodiments of the present disclosure, after the first initial risk assessment result and the second initial risk assessment result are determined, the two may be integrated to generate a risk assessment report.

[0143] Among them, the risk assessment report may include: risk score, risk level, and risk assessment.

[0144] Specifically, it includes: using text-image alignment technology to establish a spatial correspondence between text risk points and visual risk elements; using multimodal fusion technology to construct a joint representation containing text and visual features for each risk point in the first initial risk assessment result and the second initial risk assessment result; using a graph convolutional network or Transformer to capture the long-distance dependencies between risk points; using a comprehensive scoring model to generate a risk score and confidence for each risk point based on the spatial correspondence, the joint representation, and the long-distance dependencies; using hierarchical clustering or topic model technology to group each risk point and generate a document-level risk summary, which includes a risk category, an overall risk score, and a list of key risk points; representing the dependencies between different risk factors based on a probabilistic graph model; implementing a feedback loop mechanism to dynamically adjust the weights of text and visual analysis based on historical accuracy, and generate a natural language explanation for each comprehensive risk assessment result; using visualization technology to generate the risk assessment report.

[0145] In some optional examples, a time series model can be used to analyze the risk evolution trend based on the risk assessment report.

[0146] Therefore, by fusing the first initial risk assessment result with the second initial risk assessment result, it is possible to make full use of information from both textual and visual dimensions to provide a more comprehensive and accurate legal risk assessment. This multimodal fusion method can not only identify risks that may be overlooked by a single modality, but also improve the reliability of the overall assessment through cross-validation. Ultimately, the comprehensive risk report generated by the system will contain rich information from textual and visual analysis, providing comprehensive decision support for legal professionals.

[0147] In some optional examples, steps of building a risk pattern library may also be performed, including: collecting and classifying common legal risk patterns, designing a formal representation of risk patterns, and implementing an automatic update mechanism for risk patterns.

[0148] And optionally, compliance recommendation generation and system output, including: building a compliance knowledge base, including: collecting and organizing the latest laws and regulations, establishing a mapping relationship between regulations and risks, and realizing an automatic update mechanism for the knowledge base; personalized compliance recommendation generation, including: retrieving relevant regulations based on risk assessment results, using a multimodal language model to generate targeted recommendations, and applying a noise-gated Bayesian network to filter unreliable recommendations; system output and interaction, including: designing a user interface to display risks and recommendations; interactive query and interpretation functions, and regular generation of risk reports and compliance audit documents.

[0149] In order to better understand and implement the legal risk warning and compliance management method in the embodiments of the present disclosure, a detailed description is given through examples.

[0150] Example 1: Contract risk identification

[0151] In this example, the system analyzes a complex business contract to identify potential legal risks.

[0152] Step 1: Data collection and preprocessing

[0153] 1.1: Receive a scanned PDF contract document.

[0154] 1.2: Convert PDF to searchable text through OCR technology while preserving the original layout information.

[0155] 1.3: Clean the text, remove irrelevant information such as headers and footers, and perform preliminary structural processing.

[0156] Step 2: Multimodal Language Model Analysis

[0157] 2.1: The multimodal language model reads the processed contract text.

[0158] 2.2: Multimodal language models identify key elements of a contract, such as the parties to the contract, the subject matter of the contract, the performance period, breach of contract clauses, etc.

[0159] 2.3: The multimodal language model conducts in-depth analysis of each key element to identify potential risk points.

[0160] Step 3: Noise Gate Bayesian Network Correction

[0161] 3.1: Input the preliminary analysis results of the multimodal language model into the noise gate Bayesian network.

[0162] 3.2: Noise-gate Bayesian network calibrates risk assessment based on prior knowledge and historical cases.

[0163] 3.3: Filter out unreliable risk points that may be caused by the “hallucination” of the multimodal language model through the noise gate Bayesian network.

[0164] Step 4: Visual information supplement

[0165] 4.1: A multi-model collaborative visual system analyzes the layout and signature area of ​​the contract.

[0166] 4.2: Identify whether there are any abnormal modification traces or suspicious signatures.

[0167] 4.3: Integrate visual analysis results with text analysis results.

[0168] Step 5: Risk Assessment and Recommendation Generation

[0169] 5.1: The system integrates text and visual analysis results to give risk scores and levels.

[0170] 5.2: Generate targeted risk mitigation recommendations based on the latest laws and regulations.

[0171] 5.3: Generate a detailed risk assessment report, including explanations and recommendations for each risk point.

[0172] Example 2: Email Compliance Monitoring

[0173] In this example, the system will monitor the company's email communications in real time to identify potential information leaks or violations.

[0174] Step 1: Data collection and preprocessing

[0175] 1.1: The system receives the email data stream from the enterprise email server in real time.

[0176] 1.2: Parse the email and separate the body, attachments and metadata.

[0177] 1.3: Conduct preliminary identification and desensitization of sensitive information in email content.

[0178] Step 2: Ternary Systolic Array SA Processing

[0179] 2.1: Convert the parsed email data into a triple format (data, timestamp, initial credibility).

[0180] 2.2: Efficient parallel processing of mail streams through systolic array architecture.

[0181] 2.3: Real-time data synchronization and filtering based on timestamp and initial credibility.

[0182] Step 3: Multimodal Language Model Analysis

[0183] 3.1: The multimodal language model receives the email data stream processed by the ternary systolic array SA.

[0184] 3.2: Conduct in-depth semantic analysis on email content to identify potential sensitive information and illegal expressions.

[0185] 3.3: Analyze the recipients, frequency, and context of emails to identify unusual communication patterns.

[0186] Step 4: Noise Gate Bayesian Network Correction

[0187] 4.1: Input the analysis results of the multimodal language model into the noise gate Bayesian network.

[0188] 4.2: Calibrate risk assessment based on the company’s communications policy and historical breach cases.

[0189] 4.3: Filter out false positives and improve the accuracy of risk identification.

[0190] Step 5: Attachment Analysis for Multi-Model Collaboration

[0191] 5.1: Identify the type and extract the content of email attachments.

[0192] 5.2: Use specialized document analysis models and image recognition models to handle different types of attachments.

[0193] 5.3: Integrate the attachment analysis results with the email body analysis results.

[0194] Step 6: Real-time risk warning and intervention

[0195] 6.1: Based on the comprehensive analysis results, each email is given a real-time risk score.

[0196] 6.2: When a high-risk email is identified, the system immediately issues an alert.

[0197] 6.3: For certain types of high-risk emails, the system can automatically block them from being sent and notify the compliance department.

[0198] Step 7: Compliance report and recommendation generation

[0199] 7.1: The system generates regular email compliance reports, including risk trend analysis.

[0200] 7.2: Provide targeted employee training recommendations based on identified risk patterns.

[0201] 7.3: Update the enterprise’s communications policy and best practice guidelines.

[0202] Example 3: Real-time voice call compliance monitoring

[0203] In this example, the system will monitor the enterprise's voice calls (such as customer service calls) in real time to ensure compliance and improve service quality.

[0204] Step 1: Audio data collection and preprocessing

[0205] 1.1: The system receives the audio stream from the enterprise telephone system in real time.

[0206] 1.2: Perform noise reduction and sound quality enhancement on the audio stream.

[0207] 1.3: Convert audio to text using real-time speech recognition (ASR) technology.

[0208] Step 2: Ternary Systolic Array SA Processing

[0209] 2.1: Organize the converted text data into triple format (text segment, timestamp, ASR confidence).

[0210] 2.2: Use systolic array architecture to implement streaming processing of speech-to-text data.

[0211] 2.3: Perform preliminary data quality control based on ASR confidence.

[0212] Step 3: Multimodal language model real-time analysis

[0213] 3.1: The multimodal language model receives the processed text stream and performs real-time semantic analysis.

[0214] 3.2: Identify keywords, sentiment changes, and potential compliance risks in conversations.

[0215] 3.3: Analyze the structure and progression of conversations to identify possible quality of service issues.

[0216] Step 4: Noise Gate Bayesian Network Application

[0217] 4.1: Input the real-time analysis results of the multimodal language model into the noise gate Bayesian network.

[0218] 4.2: Conduct credibility assessment on identified risk points based on industry standards and corporate policies.

[0219] 4.3: Filter out false positive results due to speech recognition errors or context misunderstandings.

[0220] Step 5: Multimodal analysis integration

[0221] 5.1: Combine the acoustic features of the audio (such as pitch, volume, speaking rate) with the text analysis results.

[0222] 5.2: Use sentiment analysis models to assess the speaker’s emotional state.

[0223] 5.3: Integrate multimodal analysis results to provide more comprehensive dialogue understanding.

[0224] Step 6: Real-time intervention and guidance

[0225] 6.1: When the system detects potential compliance risks or service quality issues, it immediately warns customer service personnel.

[0226] 6.2: Provide customer service staff with advice on how to speak and compliance guidance based on real-time analysis results.

[0227] 6.3: For serious violation risks, the system can automatically notify supervisors for real-time intervention.

[0228] Step 7: Post-processing analysis and improvement

[0229] 7.1: Conduct a comprehensive post-event analysis of each call and generate detailed compliance reports.

[0230] 7.2: Based on the accumulated analysis results, continuously update and optimize the system's risk identification model.

[0231] 7.3: Generate personalized training recommendations to help customer service staff improve compliance awareness and service skills.

[0232] Therefore, through the above examples, the application capabilities of this technical solution in different scenarios are demonstrated. By combining innovative technologies such as multimodal large language models, noise gate Bayesian networks, ternary systolic array SA and multi-model collaborative visual information extraction, this solution can provide efficient and accurate legal risk warning and compliance management services in various complex business environments. This not only greatly improves the risk management capabilities of enterprises, but also provides strong support for improving overall operational efficiency and service quality.

[0233] The disclosed embodiment also provides a system corresponding to legal risk warning and compliance management, such as Figure 4 The structure diagram of a legal risk warning and compliance management system in an embodiment of the present disclosure is shown in FIG. Figure 4 As shown, the legal risk warning and compliance management system 100 may include:

[0234] The data acquisition module 110 is adapted to acquire initial data related to the law and the contract, and pre-process the initial data to generate structured analysis data, wherein the initial data includes at least one of text data, image data, and audio data;

[0235] The semantic analysis module 120 is adapted to construct a ternary metadata structure of the analysis data, and perform semantic analysis on the ternary metadata structure of the initial data, identify data with potential risk points in the analysis data, and give corresponding risk scores as target data;

[0236] The inference module 130 is adapted to output the target data to a pre-trained Bayesian network, perform inference processing on the target data, and generate a first initial risk assessment result; and is adapted to perform visual analysis processing on the initial data, and generate a second initial risk assessment result;

[0237] The generating module 140 is adapted to generate a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result.

[0238] Among them, the specific working process and principle of the data acquisition module 110, the semantic analysis module 120, the inference module 130 and the generation module 140 can refer to the relevant description of the above examples.

[0239] In short, this solution proposes an intelligent legal risk warning and compliance management system based on a multimodal large language model. Through innovative model architecture and algorithm design, it realizes real-time analysis of various types of enterprise data, risk identification and compliance recommendation generation, so as to improve the enterprise's risk management capabilities and provide strong support for improving overall operational efficiency and service quality.

[0240] The present disclosure also provides a computer system suitable for implementing a legal risk early warning and compliance management method, wherein the computer system includes a central processing unit (CPU), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage part into a random access memory (RAM), such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM. The CPU, ROM, and RAM are connected to each other via a bus. The input / output (I / O) interface is also connected to the bus.

[0241] The following components are connected to the I / O interface: an input part including a keyboard, a mouse, etc.; an output part including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage part including a hard disk, etc.; and a communication part including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication part performs communication processing via a network such as the Internet. The drive 310 is also connected to the I / O interface as needed. Removable media, such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive as needed so that the computer program read therefrom can be installed into the storage part as needed.

[0242] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication part, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), various functions defined in the system of the present application are executed.

[0243] It should be noted that the computer-readable medium shown in the embodiment of the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0244] In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable computer program. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0245] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0246] The units involved in the embodiments described in the present disclosure may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not constitute limitations on the units themselves in some cases.

[0247] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the above-mentioned various optional implementations.

[0248] As another aspect, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

[0249] It should be noted that, although several modules or units of the equipment for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into being embodied by multiple modules or units.

[0250] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the implementation method of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation method of the present application.

[0251] Although the embodiments of the present disclosure are disclosed as above, the present invention is not limited thereto. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A legal risk early warning and compliance management method, It is characterized in that include: Acquire initial data related to the law and the contract, and pre-process the initial data to generate structured analysis data, wherein the initial data includes at least one of text data, image data, and audio data; Constructing a ternary metadata structure of the analysis data, and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data; Outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result; Performing visual analysis processing on the initial data to generate a second initial risk assessment result, including: using a pre-trained YOLOv5 model to identify structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements; using a ResNet model to extract the initial data for semantic analysis, and extracting key information in the image area corresponding to the initial data; using an attention mechanism to fuse the logical structure tree and the key information, and based on the confidence scores of the logical structure tree and the key information, adjusting the fusion ratio of the logical structure tree and the key information to obtain visual information corresponding to the initial data; converting the visual information to obtain a sequence representation of the visual information; performing a structured analysis on the sequence representation of the visual information to identify the risk points and confidence levels of the sequence list; and generating a second initial risk assessment result based on the risk points and confidence levels of the sequence list; generating a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result; The step of constructing the ternary metadata structure of the analysis data and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data includes: Generating a ternary systolic array includes: defining a circular buffer having a fixed size, and defining ternary data in the circular buffer; Adopting a cyclic writing strategy and a time window mechanism, the analysis data acquired within a specific time is input into the ternary systolic array, and a ternary data structure corresponding to the type of the analysis data is generated, wherein the ternary data structure includes: data, timestamp and credibility; Inputting the ternary data structures into a pre-trained multimodal language model in batches, performing semantic analysis on the ternary data structures of the analysis data through the multimodal language model, identifying key elements in the ternary data structures of the analysis data, analyzing the key elements, determining data with potential risk points in the analysis data, and giving corresponding risk scores; The Bayesian network includes: input nodes, hidden nodes and output nodes, wherein the hidden nodes are provided with conditional probability tables generated based on the empirical coding of the target population and evaluation data sets for characterizing various legal scenarios and different types of contracts; Outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result includes: The target data is discretized through the input node to obtain discrete data, and output to the hidden node; By using the hidden nodes, based on the discrete data, the conditional probability table and the evaluation data set, an expectation maximization algorithm is applied to generate the first initial risk assessment result; and outputting the first initial risk assessment result through the output node; The step of generating a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result includes: Use text-image alignment technology to establish the spatial correspondence between text risk points and visual risk elements; Using multimodal fusion technology, constructing a joint representation including text and visual features for each risk point in the first initial risk assessment result and the second initial risk assessment result; Use graph convolutional networks or Transformers to capture long-distance dependencies between risk points; Using a comprehensive scoring model, based on the spatial correspondence, the joint representation, and the long-distance dependency, to generate a risk score and confidence level for each risk point; Use hierarchical clustering or topic modeling techniques to group risk points and generate a document-level risk summary, which includes a risk category, an overall risk score, and a list of key risk points; Represent the dependencies between different risk factors based on probabilistic graphical models; Implement a feedback loop mechanism to dynamically adjust the weight of text and visual analysis based on historical accuracy and generate natural language explanations for each integrated risk assessment result; Using visualization technology, the risk assessment report is generated.

2. According to the legal risk early warning and compliance management method of claim 1, It is characterized in that The step of adopting a cyclic writing strategy and a time window mechanism to input the acquired analysis data within a specific time into the ternary systolic array further includes: A weighted average or aggregation algorithm is used to determine the credibility of the analysis data, and analysis data with a credibility lower than a set credibility is filtered out.

3. According to the legal risk early warning and compliance management method of claim 1, It is characterized in that The multimodal language model is obtained in the following manner: Selecting a multi-objective transformer model, integrating a multimodal processing module on the multi-objective transformer model, and configuring an attention mechanism on the multi-objective transformer model; Collect legal text corpus and contract text forecasts, as well as pre-constructed legal risk identification statements and compliance assessment data sets, train the multi-objective transformer model, and generate the multimodal language model.

4. According to the legal risk early warning and compliance management method of claim 2, It is characterized in that In the case where the initial data is audio data and / or video data, after converting the audio data and / or video data into corresponding text data, after obtaining the converted text data, the ternary data structure of the converted text data is constructed using the ternary systolic array; Wherein, when the initial data is audio data, the ternary data structure includes: a text segment, a timestamp, and an ASR confidence level; When the initial data is image data, the ternary data structure includes: a text segment, a timestamp, and an OCR confidence level.

5. According to the legal risk early warning and compliance management method of claim 1, It is characterized in that The step of discretizing the target data through the input node to obtain discrete data further includes: A noise gate mechanism is set on the input node, and target data having a risk score lower than a preset risk score corresponding to the noise gate mechanism is filtered out.

6. A legal risk warning and compliance management system, It is characterized in that include: A data acquisition module, adapted to acquire initial data related to laws and contracts, and pre-process the initial data to generate structured analysis data, wherein the initial data includes at least one of text data, image data, and audio data; A semantic analysis module, adapted to construct a ternary metadata structure of the analysis data, and perform semantic analysis on the ternary metadata structure of the initial data, identify data with potential risk points in the analysis data, and give corresponding risk scores as target data; An inference module, adapted to output the target data to a pre-trained Bayesian network, perform inference processing on the target data, and generate a first initial risk assessment result; and adapted to perform visual analysis processing on the initial data to generate a second initial risk assessment result, comprising: using a pre-trained YOLOv5 model to identify structural elements in the initial data, and constructing a logical structure tree for characterizing the hierarchical relationship between the elements; using a ResNet model to extract the initial data for semantic analysis, and extracting key information in the image area corresponding to the initial data; using an attention mechanism to fuse the logical structure tree and the key information, and based on the confidence scores of the logical structure tree and the key information, adjusting the fusion ratio of the logical structure tree and the key information to obtain visual information corresponding to the initial data; converting the visual information to obtain a sequence representation of the visual information; performing a structured analysis on the sequence representation of the visual information to identify the risk points and confidence levels of the sequence table; and generating a second initial risk assessment result based on the risk points and confidence levels of the sequence table; A generating module, adapted to generate a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result; The step of constructing the ternary metadata structure of the analysis data and performing semantic analysis on the ternary metadata structure of the initial data, identifying data with potential risk points in the analysis data, and giving corresponding risk scores as target data includes: Generating a ternary systolic array includes: defining a circular buffer having a fixed size, and defining ternary data in the circular buffer; Adopting a cyclic writing strategy and a time window mechanism, the analysis data acquired within a specific time is input into the ternary systolic array, and a ternary data structure corresponding to the type of the analysis data is generated, wherein the ternary data structure includes: data, timestamp and credibility; Inputting the ternary data structures into a pre-trained multimodal language model in batches, performing semantic analysis on the ternary data structures of the analysis data through the multimodal language model, identifying key elements in the ternary data structures of the analysis data, analyzing the key elements, determining data with potential risk points in the analysis data, and giving corresponding risk scores; The Bayesian network includes: input nodes, hidden nodes and output nodes, wherein the hidden nodes are provided with conditional probability tables generated based on the empirical coding of the target population and evaluation data sets for characterizing various legal scenarios and different types of contracts; Outputting the target data to a pre-trained Bayesian network, performing inference processing on the target data, and generating a first initial risk assessment result includes: The target data is discretized through the input node to obtain discrete data, and output to the hidden node; By using the hidden nodes, based on the discrete data, the conditional probability table and the evaluation data set, an expectation maximization algorithm is applied to generate the first initial risk assessment result; and outputting the first initial risk assessment result through the output node; The step of generating a risk assessment report according to the first initial risk assessment result and the second initial risk assessment result includes: Use text-image alignment technology to establish the spatial correspondence between text risk points and visual risk elements; Using multimodal fusion technology, constructing a joint representation including text and visual features for each risk point in the first initial risk assessment result and the second initial risk assessment result; Use graph convolutional networks or Transformers to capture long-distance dependencies between risk points; Using a comprehensive scoring model, based on the spatial correspondence, the joint representation, and the long-distance dependency, to generate a risk score and confidence level for each risk point; Use hierarchical clustering or topic modeling techniques to group risk points and generate a document-level risk summary, which includes a risk category, an overall risk score, and a list of key risk points; Represent the dependencies between different risk factors based on probabilistic graphical models; Implement a feedback loop mechanism to dynamically adjust the weight of text and visual analysis based on historical accuracy and generate natural language explanations for each integrated risk assessment result; Using visualization technology, the risk assessment report is generated.

Images