Multi-modal dynamic isolated forest contract risk analysis method, system, equipment and medium

Abstract

The embodiment of the invention provides a multi-modal dynamic isolated forest contract risk analysis method, system and device and a medium, and belongs to the field of contract risk analysis. The method comprises the steps that multi-modal data related to contract risks are collected, feature extraction is carried out, and feature vectors corresponding to the multi-modal data are obtained; constructing a dynamic isolated forest; inputting the feature vector corresponding to the multi-modal data into a dynamic isolated forest, and determining a multi-modal abnormal score by integrating path lengths of all trees in the dynamic isolated forest so as to reflect the abnormal degree of the multi-modal data; and according to the multi-modal anomaly score, extracting a division path causing the anomaly, and inputting the division path and the contract original text into a large language model to generate a risk report. The semantic understanding ability of the large model is deeply combined with the anomaly detection structure of the isolated forest, the risk analysis efficiency and precision are improved, the large model is used as a multi-modal aligner, and the cross-modal detection accuracy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of contract risk analysis, and particularly to a method, system, device and medium for analyzing multi-modal dynamic isolated forest contract risks. Background Art

[0002] With the increase in the complexity of enterprise business and the improvement of the requirements for compliance and efficiency, intelligent contract management systems have become one of the core systems indispensable to many enterprises. The intelligent application of intelligent contract management systems can improve the efficiency and accuracy of enterprise contract management, solve the pain points in traditional contract management, reduce contract risks, strengthen the control of contracts and realize the sharing of contract information, providing better support for enterprise contract management.

[0003] Contract risk management is crucial in fields such as finance, law, and supply chain, but traditional technologies have the following problems: 1. Semantic blind spot: Existing algorithms, such as rule engines, can only process structured data and cannot understand the implicit risks in specific contract terms, such as vague expressions like "can be terminated unilaterally". Especially when there are a large number of contract terms, traditional solutions are inefficient and inaccurate.

[0004] 2. Multi-modal fragmentation: Risk factors are often scattered in text (terms), numerical values (payment records), and graphs (enterprise associations). Existing tools lack the ability to perform cross-modal correlation analysis.

[0005] 3. Insufficient interpretability: Regulatory requirements provide legal basis for risk decisions, but traditional anomaly detection only outputs scores, which require manual interpretation by experts, etc., and is inefficient. Summary of the Invention

[0006] The purpose of the embodiments of the present invention is to provide a method, system, device and medium for analyzing multi-modal dynamic isolated forest contract risks, which deeply combines the semantic understanding ability of large models with the anomaly detection structure of isolated forests, improves the efficiency and accuracy of risk analysis, uses large models as multi-modal aligners, greatly improves the cross-modal detection accuracy, and uses large models to automatically generate analysis reports to improve the interpretation efficiency.

[0007] To achieve the above purpose, the embodiments of the present invention provide a method for analyzing multi-modal dynamic isolated forest contract risks, including: Collect multi-modal data related to contract risks and perform feature extraction to obtain feature vectors corresponding to the multi-modal data; Construct a dynamic isolated forest; Input the feature vectors corresponding to the multi-modal data into the dynamic isolated forest, and determine the multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolated forest to reflect the anomaly degree of the multi-modal data; Extract the division path that causes the anomaly based on the multimodal anomaly score, and input the division path and the original contract text into the large language model to generate a risk report.

[0008] Optionally, collect multimodal data related to contract risks and perform feature extraction to obtain feature vectors corresponding to the multimodal data, including: Collect text data, numerical data, and graph data related to contract risks; For text data, use the large language model to extract text features, convert the contract clause text into a digital vector, and use the digital vector as the feature vector; For numerical data, scale the numerical data using the median and the interquartile range of the data to obtain the feature vector corresponding to the numerical data; For graph data, use the graph neural network algorithm to generate the node representation of the enterprise association graph, and use the node representation as the feature vector corresponding to the graph data.

[0009] Optionally, convert text data into a digital vector according to the following formula:

[0010] In the formula, LLM represents the large language model, It means that the generated digital vector is a real vector with a dimension of 768.

[0011] Optionally, scale the numerical data using the median and the interquartile range of the data according to the following formula to obtain the feature vector corresponding to the numerical data:

[0012] In the formula, median(X) represents the median of all historical data, IQR(X) is the interquartile range of the data, and x is the numerical data.

[0013] Optionally, for graph data, generate the node representation of the enterprise association graph according to the following formula to obtain the feature vector corresponding to the graph data:

[0014] In the formula, h graph is the important information and attributes of the node in the network, GraphSAGE is a graph neural network algorithm for learning the representation of nodes from a graph, represents an enterprise node in the enterprise association graph, represents the entire graph structure including all nodes and the connection relationships between nodes, It means that the generated node feature vector is a real vector with a dimension of 256.

[0015] Optionally, the construction process of the dynamic isolation forest includes: Input the contract clause text into the large language model to generate a dynamic weight vector; According to the dynamic weight vector, assign weights to each feature, and reselect feature data according to the weights for splitting the isolation tree; According to the modal weighted calculation method, calculate the path length of the multi-modal data related to the contract risk in the t-th isolation tree, so as to synthesize the abnormal signals of text data, numerical data and graph data, and realize the cross-modal joint judgment of the dynamic isolation forest.

[0016] Optionally, calculate the multi-modal anomaly score according to the following formula to reflect the anomaly degree of the multi-modal data: ; ; In the formula, E[h multi (x)] represents the average path length of the multi-modal samples in all trees, c(W) is the standardized reference value dynamically adjusted according to the feature weights, H(W) is the entropy of the weight W distribution, n represents the number of samples, W i is the weight of feature i, and d represents the total number of features.

[0017] In the second aspect, the present invention also provides an analysis system for multi-modal dynamic isolation forest contract risk, including: An acquisition unit for acquiring multi-modal data related to the contract risk and performing feature extraction to obtain a feature vector corresponding to the multi-modal data; A construction unit for constructing a dynamic isolation forest; An evaluation unit for inputting the feature vector corresponding to the multi-modal data into the dynamic isolation forest, and determining the multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolation forest to reflect the anomaly degree of the multi-modal data; An analysis unit for extracting the division path causing the anomaly according to the multi-modal anomaly score, and inputting the division path and the contract original text into the large language model to generate a risk report.

[0018] In the third aspect, an embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned analysis method for multi-modal dynamic isolation forest contract risk are implemented.

[0019] In the fourth aspect, the present invention also provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned analysis method for multi-modal dynamic isolation forest contract risk are implemented.

[0020] Through the above technical solutions, the semantic understanding ability of the large model is deeply combined with the anomaly detection structure of the Isolation Forest, improving the efficiency and accuracy of risk analysis. Using the large model as a multi-modal aligner can greatly improve the cross-modal detection accuracy compared with traditional methods. Using the large model to automatically generate an analysis report containing legal basis can improve the interpretation efficiency.

[0021] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent specific implementation part. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings are used to provide a further understanding of the embodiments of the present invention, and constitute a part of the specification. Together with the following specific implementation, they are used to explain the embodiments of the present invention, but do not constitute a limitation to the embodiments of the present invention. In the drawings: Figure 1 is a flowchart of the implementation of an analysis method for multi-modal dynamic Isolation Forest contract risks provided by an embodiment of the present invention; Figure 2 is a flowchart of the construction process of a dynamic Isolation Forest provided by an embodiment of the present invention; Figure 3 is a detailed flowchart of the implementation of an analysis method for multi-modal dynamic Isolation Forest contract risks provided by an embodiment of the present invention; Figure 4 is a schematic structural diagram of an analysis system for multi-modal dynamic Isolation Forest contract risks provided by an embodiment of the present invention; Figure 5 is a schematic hardware structure diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] In the following detailed description, various embodiments of the present disclosure will be described more fully. The present disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the present disclosure to the specific embodiments disclosed herein, but rather the present disclosure should be understood to cover all adjustments, equivalents, and / or alternative solutions falling within the spirit and scope of the various embodiments of the present disclosure.

[0024] In the following, the term "comprise" or "may comprise" that may be used in various embodiments of the present disclosure indicates the presence of the disclosed functions or operations, and does not limit the addition of one or more functions or operations. Further, as used in various embodiments of the present disclosure, the terms "comprise", "have" and their cognates are only intended to indicate specific features, numbers, steps, operations or combinations of the foregoing items, and should not be construed as precluding the existence or addition of one or more other features, numbers, steps, operations or combinations of the foregoing items.

[0025] In various embodiments of the present disclosure, the expression "or" or "at least one of A or / and B" includes any combination or all combinations of the recited words. For example, the expression "A or B" or "at least one of A or / and B" may include A, may include B, or may include both A and B.

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0027] Refer to Figure 1 Shown is a flowchart of an analysis method for multi-modal dynamic isolated forest contract risks in a specific embodiment, including the following execution steps: Step 100: Collect multi-modal data related to contract risks and perform feature extraction to obtain a feature vector corresponding to the multi-modal data.

[0028] In one embodiment of the present invention, based on Step 100, a possible embodiment will be given below to non-restrictively elaborate on its specific implementation scheme. It includes the following execution steps: S1000: Collect text data, numerical data, and graph data related to contract risks.

[0029] Exemplarily, the text data includes, but is not limited to, contract terms, supplementary agreements, and historical dispute records; the numerical data includes, but is not limited to, payment amounts, performance times, and liquidated damages ratios; the graph data includes, but is not limited to, enterprise equity relationships and signatory association networks.

[0030] S1001: For the text data, use a large language model to extract text features, convert the contract clause text into a digital vector, and use the digital vector as the feature vector.

[0031] For text data, use Legal-BERT or other fine-tuned LLMs for the legal domain to extract text features, converting the text of contract terms into digital vectors that can be processed by a computer. Specifically, the text data can be converted into digital vectors according to the following formula:

[0032] In the formula, LLM represents a large language model, indicating that the generated digital vector is a real number vector with a dimension of 768.

[0033] S1002: For numerical data, use the median and the interquartile range of the data to scale the numerical data and obtain the feature vector corresponding to the numerical data.

[0034] In a specific embodiment, for numerical data, eliminate the influence of extreme values of numerical values such as amounts and times to avoid distorting the overall risk judgment due to a single extremely large amount, illegal time format, etc. Use the RobustScaling method. Specifically, the median and the interquartile range of the data can be used to scale the numerical data according to the following formula to obtain the feature vector corresponding to the numerical data:

[0035] In the formula, median(X) represents the median of all historical data, IQR(X) is the interquartile range of the data, and x is the numerical data.

[0036] Exemplarily, assume that the median of historical payment amounts is 1 million and the IQR is 500,000. Then a payment of 2 million is standardized to: (2 - 1) / 0.5 = 2.0, meaning that values exceeding 2.0 will be regarded as abnormal candidates).

[0037] S1003: For graph data, use a graph neural network algorithm to generate node representations of the enterprise association graph and use the node representations as the feature vectors corresponding to the graph data.

[0038] In a specific implementation manner, for graph data, use the GraphSAGE algorithm (a graph neural network algorithm that generates feature vectors of nodes by aggregating information of nodes and their neighbors) to generate node representations of the enterprise association graph. Specifically, generate node representations of the enterprise association graph according to the following formula to obtain the feature vectors corresponding to the graph data:

[0039] In the formula, h graphIt is the generated node feature vector, representing the features of a node in the graph, which is used to capture the important information and attributes of the node in the network. GraphSAGE is a graph neural network algorithm for learning node representations from graphs. It represents an enterprise node in the enterprise association graph. It represents the entire graph structure that includes all nodes and the connection relationships between nodes. It represents that the generated node feature vector is a real vector with a dimension of 256.

[0040] Exemplarily, assume there is an enterprise association graph. Each node in the graph represents an enterprise, and the edges represent certain associations (such as cooperation, investment, etc.) between these enterprises. The goal is to generate a feature vector for each enterprise, which can capture the key information of the enterprise and its position in the network. By analyzing these feature vectors, potential risks can be identified.

[0041] The graph neural network algorithm can automatically learn the representation of enterprises without manually designing feature extractors. The generated node representations (feature vectors) can capture the complex association relationships between enterprises. Using the feature vectors generated by the graph neural network as input can improve the performance of downstream tasks (such as enterprise risk prediction, recommendation systems, etc.). The feature vectors can reflect the position and role of enterprises in the graph, supporting fine-grained enterprise analysis. By analyzing the feature vectors and the association relationships between nodes, the interpretability of the model can be enhanced.

[0042] Step 101: Construct a dynamic isolation forest.

[0043] In an embodiment of the present invention, referring to Figure 2 as shown, based on Step 101, the following will give a possible embodiment to non-restrictively elaborate on its specific implementation scheme, including the following execution steps: S1010: Input the contract clause text into the large language model to generate a dynamic weight vector.

[0044] Specifically, inputting the contract clause text into the large language model to generate a dynamic weight vector can be achieved according to the following formula:

[0045] Among them, the purpose of the LLM output function is that the LLM returns the initial risk score for each clause text (for example, the contract price clause returns 0.9 points). The Softmax function compresses all scores to between 0 and 1 and the sum is 1 (similar to "allocating attention weights"). V is a trainable projection matrix used to adapt to different contract types.

[0046] Thus, the high-risk points can be actively focused on through the semantic understanding of the LLM.

[0047] S1011: Assign weights to each feature according to the dynamic weight vector, and re-select feature data according to the weights for the splitting of the isolation tree.

[0048] Specifically, assign a weight to each feature, and then select features according to these weights:

[0049] where P(selecting feature i) represents the probability of selecting feature i. Simply put, it is the probability of selecting a specific feature for splitting, and W i is the weight of feature i (the weight is used to reflect the importance or relevance of this feature), and ∑dj = 1W j refers to the sum of all feature weights, and d represents the total number of features.

[0050] Exemplarily, assume there are three features: text feature T, numerical feature N, and atlas feature G, and the corresponding weights are wT = 0.4, wN = 0.3, and wG = 0.3 respectively. When selecting the splitting feature, first calculate the weighted random value of each feature. For example, if a certain value of text feature T is t, then the weighted random value is 0.4t; if a certain value of numerical feature N is n, then the weighted random value is 0.3n; if a certain value of atlas feature G is g (here g can be a certain quantified atlas index), then the weighted random value is 0.3g, and select the feature with the largest weighted random value for splitting.

[0051] By selecting features for splitting according to the weight vector, the isolation tree can pay more attention to the features that have a greater impact on the contract risk, thereby improving the accuracy of anomaly detection; the dynamic weight vector can be adjusted according to different contract risk scenarios and data distributions, enabling the model to adapt to different environments and requirements. By analyzing the weight vector, it can be understood which features have a greater impact on the contract risk, so as to conduct a more fine-grained risk analysis.

[0052] S1012: Calculate the path length of the multi-modal data related to the contract risk in the t-th isolation tree according to the modal weighted calculation method, so as to synthesize the anomaly signals of the text data, numerical data, and graph data, and realize the cross-modal joint judgment of the dynamic isolation forest.

[0053] By weighted calculation of abnormal signals in different modalities (text, numerical values, graphs), the dynamic isolation forest can comprehensively consider the information of each modality, avoid insufficient or misleading information in a single modality, and thus improve the accuracy of anomaly detection. The weights can be dynamically adjusted according to the importance and relevance of different modality data. For example, in some contract risk scenarios, text data may be more critical, while in other scenarios, numerical data or graph data may be more important. Dynamically adjusting the weights enables the model to adapt to different contract risk scenarios. Through weighted calculation, the model can achieve cross-modal joint judgment instead of simply splicing or averaging the results of each modality. This joint judgment method can more accurately capture the complex patterns and associations in contract risks. Moreover, through weighted calculation, the model can not only give an overall anomaly score but also analyze the contribution of each modality to the anomaly score, which helps users understand the specific sources and nature of contract risks in more fine-grained detail.

[0054] Risk tracing and explanation: When an anomaly is detected, users can trace the source and cause of the risk by analyzing the weights and contributions of each modality, providing strong support for risk response.

[0055] Specifically, to integrate abnormal signals of various data types such as text, numerical values, and graphs, for a sample x, the path length h t (x) of the sample in the t-th tree is weighted by modality:

[0056] where h(k) t(x) refers to the abnormal path length of the sample in the k-th modality (such as text or numerical value), and λ k is the modality importance coefficient (dynamically calculated through the attention mechanism).

[0057] Exemplarily, cross-modal joint judgment can be analogously understood as follows: When a judge tries a case, the judge not only looks at the contract text (text), checks the transfer records (numerical values), but also investigates the enterprise relationship (graph), and finally assigns credibility to different evidences.

[0058] Step 102: Input the feature vectors corresponding to the multi-modal data into the dynamic isolation forest, and determine the multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolation forest to reflect the anomaly degree of the multi-modal data.

[0059] Multimodal data usually contains data from different sources or different types (such as images, texts, audios, etc.). The dynamic isolation forest can handle high-dimensional feature vectors, effectively integrate multimodal information, and thus detect anomalies more comprehensively. By integrating the path lengths of all trees in the dynamic isolation forest, the anomaly score of each data point can be calculated. The shorter the path length, the easier it is to isolate the data point, and the higher the degree of anomaly. The anomaly score provides a quantitative index that can intuitively reflect the anomaly degree of multimodal data, facilitating subsequent analysis and processing.

[0060] Specifically, the multimodal anomaly score is calculated according to the following formula to reflect the anomaly degree of multimodal data: ; The function of this formula is to calculate the "anomaly degree" of a certain contract clause or transaction record. The closer the score is to 1, the higher the risk; the closer it is to 0, the more normal. E[hmulti(x)] represents the average path length of the sample in all trees (which can be understood as "the number of steps required to detect an anomaly"), and c(W) is a standardized reference value dynamically adjusted according to the feature weights. It is transformed into: ; In the formula, E[h multi (x)] represents the average path length of the multimodal sample in all trees, c(W) is a standardized reference value dynamically adjusted according to the feature weights, H(W) is the entropy of the weight W distribution, used to measure uncertainty. The more concentrated the weights (such as the weight of a certain clause is 0.9), the lower the entropy. For example, input the weights W1, W2, W3 of all clauses (arbitration clause = 0.8, payment method = 0.1, others = 0.1), then H(W)=−(0.8log0.8 + 0.1log0.1 + 0.1log0.1). n represents the number of samples. As the sample size n increases, the influence of the correction term becomes smaller (more stable under big data). W i is the weight of feature i, and d represents the total number of features.

[0061] In some embodiments, when analyzing contract risks, it is necessary to flexibly grasp the risk scale according to different contract types. The implementation plan is: first, let the large model analyze the contract type and then give adjustment suggestions, and then let the system automatically adjust the risk warning line according to different contract types:

[0062] Among them, θ base is the basic threshold, representing the initial risk standard common to all contracts (such as 0.5 points), and Δ represents the maximum adjustment amplitude preset by the system (such as set to 0.3). For example, the basic threshold θ baseIf μ = 0.5, the maximum adjustment amplitude Δ = 0.3, and the current contract LLM output = +0.8, then θ = 0.5 + 0.8 × 0.3 = 0.74.

[0063] In this way, according to the characteristics and risk levels of different contract types, the risk warning line is automatically adjusted to ensure the pertinence and effectiveness of risk monitoring. As the contract execution process progresses, the system can update the risk status in real time and dynamically adjust the risk warning line according to the new risk situation. Moreover, when the risk indicators approach or exceed the warning line, the system can issue a warning in a timely manner to remind relevant personnel to take countermeasures.

[0064] Step 103: Extract the division path that causes the anomaly based on the multimodal anomaly score, and input the division path and the contract original text into the large language model to generate a risk report.

[0065] Specifically, when executing step 103, the following steps can be specifically executed: S1030: Set the modality-specific threshold according to the multimodal anomaly score, and filter out the abnormal samples whose scores exceed the modality-specific threshold.

[0066] Exemplarily, based on the anomaly score distribution of the full sample, select the quantile threshold, or set independent thresholds for different modalities respectively.

[0067] S1031: Traverse the node splitting paths of all trees in the dynamic isolation forest, and for the abnormal samples, extract the splitting conditions from the root node to the leaf node in each tree, where the splitting conditions include feature dimensions, division thresholds, and division directions.

[0068] S1032: Align the splitting paths corresponding to different modalities, count the high-frequency cross-modal splitting rules, and quantify the feature importance by the proportion of path length to generate a joint abnormal pattern.

[0069] Exemplarily, the alignment method is to encode the text modality using keyword trigger rules, map the graph data modality using visual feature space, encode the numerical modality using threshold intervals, and perform cross-modal feature matching through a shared embedding space.

[0070] S1033: Convert the joint abnormal pattern into natural language rules, and perform visual output based on the decision chain graph and the heat map, where the decision chain graph is used to display the splitting path of the abnormal sample in the tree, and the heat map is used to mark the contribution degree of multimodal features to the anomaly score.

[0071] Exemplarily, the generation of natural language rules includes associating the splitting conditions with the legal knowledge graph and annotating the contract legal clause items violated by the rules.

[0072] Exemplarily, extract the partitioning path that leads to an anomaly: path=[{"feature": "Payment amount", "split_val": "1 million", "direction": ">"}, {"feature": "Arbitration clause", "split_val": "Exists", "direction": "="}] Input the path and the original contract text into the LLM to generate a compliance explanation: prompt=f""" Contract clause: {clause} Anomaly path: {path} Relevant regulations: Article X of "XX" Please generate a risk report: """ report=GPT-4(prompt) Output example: "High risk detected: The payment amount exceeds 1 million and the contract contains an arbitration clause, which is inconsistent with the historical transaction pattern of the signing party (refer to Case No. XX of the Final Judgment of the Shanghai High People's Court in 2023)".

[0073] By combining the large language model (LLM) with the Isolation Forest algorithm, an intelligent contract risk analysis system with semantic understanding, dynamic weight allocation, and multimodal fusion capabilities is developed. Its core breakthroughs include: real-time parsing of the semantics of contract clauses by the LLM to generate dynamic feature weights (such as increasing the weight of the "Arbitration clause" to 0.8), transforming the Isolation Forest to achieve cross-modal anomaly detection, proposing an adaptive risk threshold based on the contract type (such as a threshold of 0.7 for international trade contracts), and automatically generating a natural language report with legal basis, which greatly improves the efficiency and accuracy of enterprises in analyzing contract risks.

[0074] In one embodiment, Figure 3 For the detailed flowchart of an analysis method for multimodal dynamic Isolation Forest contract risks provided by an embodiment of the present invention, this embodiment is further optimized and extended on the basis of the above embodiments.

[0075] S300: Collect text data, numerical data, and graph data related to contract risks.

[0076] S301: For the text data, use the large language model to extract text features, convert the contract clause text into digital vectors, and use the digital vectors as feature vectors.

[0077] S302: For the numerical data, scale the numerical data using the median and the interquartile range of the data to obtain the feature vector corresponding to the numerical data.

[0078] S303: For the graph data, use the graph neural network algorithm to generate the node representations of the enterprise association graph, and use the node representations as the feature vectors corresponding to the graph data.

[0079] S304: Input the contract clause text into the large language model to generate a dynamic weight vector.

[0080] S305: According to the dynamic weight vector, assign weights to each feature, and reselect the feature data according to the weights for the splitting of the isolation tree.

[0081] S306: According to the modal weighted calculation method, calculate the path length of the multi-modal data related to the contract risk in the t-th isolation tree, so as to integrate the abnormal signals of the text data, numerical data and graph data, and realize the cross-modal joint judgment of the dynamic isolation forest.

[0082] S307: Input the feature vectors corresponding to the multi-modal data into the dynamic isolation forest, and determine the multi-modal anomaly score by integrating the path lengths of all the trees in the dynamic isolation forest, so as to reflect the abnormal degree of the multi-modal data.

[0083] S308: Extract the division path leading to the anomaly according to the multi-modal anomaly score, and input the division path and the original contract text into the large language model to generate a risk report.

[0084] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0085] Through the above technical solutions, the semantic understanding ability of the large model is deeply combined with the abnormal detection structure of the isolation forest, improving the risk analysis efficiency and accuracy. Using the large model as a multi-modal aligner greatly improves the cross-modal detection accuracy compared with traditional methods. Using the large model to automatically generate an analysis report containing legal basis improves the interpretation efficiency.

[0086] As Figure 4 shown, the following is an embodiment of the multi-modal dynamic isolation forest contract risk analysis system provided by the present disclosure. It belongs to the same inventive concept as the multi-modal dynamic isolation forest contract risk analysis method in the above embodiments. For the details not described in detail in the embodiment of the multi-modal dynamic isolation forest contract risk analysis system, reference can be made to the embodiment of the multi-modal dynamic isolation forest contract risk analysis method.

[0087] An acquisition unit, configured to acquire multi-modal data related to contract risk and perform feature extraction to obtain feature vectors corresponding to the multi-modal data; A building unit for building a dynamic isolation forest; An evaluation unit for inputting the feature vector corresponding to the multimodal data into the dynamic isolation forest, and determining a multimodal anomaly score by integrating the path lengths of all the trees in the dynamic isolation forest to reflect the anomaly degree of the multimodal data; An analysis unit for extracting the division path causing the anomaly according to the multimodal anomaly score, and inputting the division path and the original contract text into a large language model to generate a risk report.

[0088] Figure 5 It is a schematic diagram of the hardware structure of an electronic device for implementing various embodiments of the present invention.

[0089] The multimodal dynamic isolation forest contract risk analysis method provided by the embodiments of the present application can be applied to an electronic device. Those skilled in the art can understand that the electronic device structure involved in the embodiments of the present invention does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine some components, or have different component arrangements. In the embodiments of the present invention, the electronic device includes, but is not limited to, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown in the figure, their connections and relationships, and their functions are only examples and are not intended to limit the implementation of the embodiments of the present application described and / or claimed herein.

[0090] The electronic device may include a processor, an external memory interface, an internal memory, a universal serial bus (USB) interface, a charging management module, a power management module, a battery, a wireless communication module, an audio module, a speaker, a microphone, a sensor module, a key, a camera, a display screen, and a SIM card interface, etc.

[0091] It can be understood that the structure schematically shown in the embodiments of the present application does not constitute a specific limitation on the electronic device. In other embodiments of the present application, the electronic device may include more or fewer components than shown in the figure, or combine some components, or split some components, or have different component arrangements. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0092] The processor may include one or more processing units. For example, the processor may include a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Among them, different processing units may be independent devices or integrated in one or more processors.

[0093] Among them, the processor may be the nerve center and command center of the electronic device. The controller may generate operation control signals according to the instruction operation code and timing signals to complete the control of fetching and executing instructions.

[0094] A memory may also be provided in the processor for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store the instructions or data that the processor has just used or recycled. If the processor needs to use the instruction or data again, it can directly call it from this memory. This avoids repeated accesses, reduces the waiting time of the processor, and thus improves the system efficiency.

[0095] The external memory interface can be used to connect an external memory card, such as a MicroSD card, to expand the storage capacity of the electronic device. The external memory card communicates with the processor through the external memory interface to achieve the data storage function. For example, files such as music and videos are saved in the external memory card.

[0096] The internal memory can be used to store computer-executable program code, and the computer-executable program code includes instructions. The processor executes various functional applications and data processing of the electronic device by running the instructions stored in the internal memory. The internal memory may include a program storage area and a data storage area. The internal memory may include a high-speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, a universal flash storage (UFS), etc.

[0097] The wireless communication function of the electronic device can be implemented through an antenna, a wireless communication module, a modem processor, a baseband processor, etc.

[0098] The wireless communication module can provide solutions for wireless communications applied to electronic devices, including wireless local area networks (WLANs) (such as wireless fidelity (Wi-Fi) networks), Bluetooth (BT), global navigation satellite systems (GNSS), frequency modulation (FM), near field communication (NFC), infrared technology (IR), etc.

[0099] The electronic device can implement audio functions through an audio module, speakers, receivers, microphones, headphone jacks, application processors, etc.

[0100] The electronic device can implement shooting functions through an ISP, camera, video codec, GPU, display screen, application processor, etc.

[0101] The electronic device can implement display functions through a GPU, display screen, application processor, etc.

[0102] The GPU is a microprocessor for image processing, connecting the display screen and the application processor. The GPU is used to perform mathematical and geometric calculations for graphics rendering. The processor may include one or more GPUs, which execute program instructions to generate or change display information.

[0103] The display screen is used to display images, videos, etc. The display screen includes a display panel.

[0104] The above-mentioned electronic device implements the collection of multi-modal data related to contract risks in the theme name of the analysis method for multi-modal dynamic isolated forest contract risks of this application, and performs feature extraction to obtain feature vectors corresponding to the multi-modal data; constructs a dynamic isolated forest; inputs the feature vectors corresponding to the multi-modal data into the dynamic isolated forest, and determines a multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolated forest to reflect the anomaly degree of the multi-modal data; extracts the division path leading to the anomaly according to the multi-modal anomaly score, and inputs the division path and the contract original text into a large language model to generate a risk report technical solution, achieving the beneficial effects of realizing the intelligence, precision, and interpretability of contract risk analysis through a closed loop of data fusion, dynamic modeling, and natural language generation.

[0105] In the storage medium provided by this application, there is a program product capable of implementing the analysis method for multi-modal dynamic isolated forest contract risks.

[0106] The analysis method for multi-modal dynamic isolated forest contract risks includes: collecting multi-modal data related to contract risks, performing feature extraction to obtain feature vectors corresponding to the multi-modal data; constructing a dynamic isolated forest; inputting the feature vectors corresponding to the multi-modal data into the dynamic isolated forest, and determining a multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolated forest to reflect the anomaly degree of the multi-modal data; extracting the division path leading to the anomaly based on the multi-modal anomaly score, and inputting the division path and the contract original text into a large language model to generate a risk report.

[0107] In some possible implementation manners, the subject name, the analysis method and system for multi-modal dynamic isolated forest contract risks of the present disclosure can be implemented in the form of a program product, which includes program code. When the program product runs on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present disclosure described in the above "Exemplary Method" section of this specification.

[0108] The storage medium of the present disclosure can adopt any combination of one or more readable media. The readable media can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or 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.

[0109] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for analyzing the contract risk of multi-modal dynamic isolation forest, characterized in that, It includes: Collect multi-modal data related to contract risks, perform feature extraction, and obtain feature vectors corresponding to the multi-modal data; Construct a dynamic isolation forest; Input the feature vectors corresponding to the multi-modal data into the dynamic isolation forest, and determine the multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolation forest to reflect the anomaly degree of the multi-modal data; Extract the division path leading to the anomaly based on the multi-modal anomaly score, and input the division path and the original contract text into a large language model to generate a risk report.

2. The analysis method for multi-modal dynamic isolated forest contract risks according to claim 1, characterized in that Collect multi-modal data related to contract risks, perform feature extraction, and obtain feature vectors corresponding to the multi-modal data, including: Collect text data, numerical data, and graph data related to contract risks; For text data, use a large language model to extract text features, convert the contract clause text into digital vectors, and use the digital vectors as feature vectors; For numerical data, scale the numerical data using the median and the interquartile range of the data to obtain the feature vectors corresponding to the numerical data; For graph data, use a graph neural network algorithm to generate node representations of the enterprise association graph, and use the node representations as the feature vectors corresponding to the graph data.

3. The analysis method for multimodal dynamic isolated forest contract risks according to claim 2, characterized in that Convert text data into digital vectors according to the following formula: In the formula, LLM represents the large language model, indicating that the generated digital vector is a real vector with a dimension of 768.

4. The analysis method for multimodal dynamic isolated forest contract risks according to claim 2, wherein Scale the numerical data using the median and the interquartile range of the data according to the following formula to obtain the feature vectors corresponding to the numerical data: In the formula, median(X) represents the median of all historical data, IQR(X) is the interquartile range of the data, and x is the numerical data.

5. The analysis method for multi-modal dynamic isolated forest contract risks according to claim 2, wherein For graph data, generate node representations of the enterprise association graph according to the following formula: where h graph is the information and attributes of the node in the network, and GraphSAGE is a graph neural network algorithm for learning node representations from graphs, represents an enterprise node in the enterprise association graph, represents the entire graph structure containing all nodes and the connection relationships between nodes, indicates that the generated node feature vector is a real vector with a dimension of 256.

6. The analysis method for multi-modal dynamic isolated forest contract risks according to claim 1, wherein, The construction process of the dynamic isolation forest includes: Input the contract clause text into a large language model to generate a dynamic weight vector; According to the dynamic weight vector, assign weights to each feature, and reselect feature data for splitting the isolation tree according to the weights; Calculate the path length of the multi-modal data related to contract risks in the t-th isolation tree according to the modal weighting calculation method to synthesize the anomaly signals of text data, numerical data, and graph data, and achieve cross-modal joint judgment of the dynamic isolation forest.

7. The analysis method for multi-modal dynamic isolated forest contract risks according to claim 1, characterized in that, Calculate the multi-modal anomaly score according to the following formula to reflect the anomaly degree of the multi-modal data: ； ； where \(E[h multi (x)]\) represents the average path length of multimodal samples in all trees, \(c(W)\) is a normalized reference value dynamically adjusted according to feature weights, \(H(W)\) is the entropy of the weight \(W\) distribution, \(n\) represents the number of samples, \(W i is the weight of feature \(i\), and \(d\) represents the total number of features.

8. A multi-modal dynamic isolation forest contract risk analysis system, characterized in that, It includes: A collection unit for collecting multi-modal data related to contract risks, performing feature extraction, and obtaining feature vectors corresponding to the multi-modal data; A construction unit for constructing a dynamic isolation forest; An evaluation unit for inputting the feature vectors corresponding to the multi-modal data into the dynamic isolation forest, and determining the multi-modal anomaly score by integrating the path lengths of all trees in the dynamic isolation forest to reflect the anomaly degree of the multi-modal data; An analysis unit for extracting the division path leading to the anomaly based on the multi-modal anomaly score, and inputting the division path and the original contract text into a large language model to generate a risk report.

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, When the processor executes the program, it implements the steps of the analysis method for multi-modal dynamic isolation forest contract risks as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for analyzing the multi-modal dynamic isolation forest contract risk according to any one of claims 1 to 7.

Images