A deep learning-based judicial multi-dimensional evaluation decision method and system

Abstract

The application discloses a kind of based on deep learning's judicial multidimensional evaluation decision-making method and system, through introducing feature vector generation, the collaborative analysis of legal application dimension and social influence dimension, judge weight correction instruction response and the difference analysis based on historical data etc. Link, a closed-loop evaluation process is constructed.The method can real-time fusion judge's professional judgment in the evaluation process, dynamically adjust model parameters, so that the final output is closer to actual judicial practice.Improve the adaptability and interactivity of judicial assistance system, so that the evaluation result can reflect the balance between legal provisions and social influence, enhance the practicability and credibility of artificial intelligence in judicial decision support.

Description

Technical Field

[0001] The present invention belongs to the field of judicial management technology, and specifically relates to a judicial multi-dimensional evaluation decision-making method and system based on deep learning. Background Art

[0002] Current judicial decision-making support systems primarily rely on statistical analysis of historical judgment data or on models based on fixed rules to evaluate cases. For example, some systems extract basic case information, such as the cause of action, the identities of the parties involved, and the type of behavior, feed this into a pre-set evaluation model, and output a reference result with a high degree of similarity to historical precedents.

[0003] However, such methods usually ignore the dynamic trade-offs between legal application and social impact made by judges in the actual process of adjudicating cases, and lack a real-time feedback mechanism for human-computer collaboration, resulting in evaluation results that are difficult to adapt to the complexity and subjective judgment factors of specific cases. Summary of the Invention

[0004] The purpose of the present invention is to provide a judicial multi-dimensional evaluation and decision-making method and system based on deep learning, which dynamically adjusts the output logic of the evaluation model according to the interactive feedback of judges during the evaluation process, realizes intelligent assisted decision-making in human-computer collaboration, thereby improving the rationality and acceptability of the evaluation results, and solving the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a judicial multi-dimensional evaluation decision-making method based on deep learning, comprising the following steps:

[0006] Receive case elements input by the judge, generate a feature vector of the case, and extract the legal application dimension and social impact dimension of the case based on the feature vector;

[0007] After collaboratively analyzing the legal applicability and social impact dimensions, an initial assessment result is generated. Deep learning technology is used to analyze historical judicial data to help judges quickly find relevant precedents. The initial assessment result is fed back to the judge's interactive interface to obtain the judge's instructions on how to adjust the weights of each dimension. Combined with natural language processing and speech recognition technology, a digital human cartoon avatar intelligent assistant provides support for case inquiries, information retrieval, and legal consultation.

[0008] According to the weight modification instructions, the calculation parameters of the legal application dimension and the social impact dimension are dynamically adjusted;

[0009] The adjusted dimension combination is recalculated to generate a revised evaluation result, and the revised evaluation result is compared with the historical judgment data to form a difference analysis report.

[0010] Preferably, generating the feature vector of the case includes:

[0011] Perform semantic analysis on the case elements input by the judge to extract the behavior types and subject relationships involved in the case;

[0012] Mapping the behavior type to a numerical identifier in a preset coding system to form a basic description vector;

[0013] The responsibility weights between the parties are determined according to the subject relationship, the associated influence factors are calculated, and the basic description vector and the associated influence factors are combined to generate the final feature vector through a weighted fusion operation.

[0014] Preferably, extracting the legal application dimension and social impact dimension of the case based on the feature vector includes:

[0015] Inputting the feature vector into a preset classification framework to identify the categories of legal provisions related to the case and form a preliminary legal label set;

[0016] Based on the preliminary legal label set, find the applicable conditions and discretionary scope of the corresponding provisions and construct the legal application dimension;

[0017] Based on the field to which the case belongs and the legal application dimensions, the related records in the social evaluation database are matched and the social impact dimensions are extracted.

[0018] Preferably, the initial assessment results are generated after a collaborative analysis based on the legal applicability dimension and the social impact dimension, including:

[0019] Determine a case discretion benchmark interval based on the legal application dimension, and perform offset correction on the discretion benchmark interval based on the social impact dimension to obtain an adjusted central value;

[0020] A floating range is set around the adjusted central value, a confidence region is generated based on the historical case distribution density, the confidence region is mapped into an interpretable scoring system, and an initial evaluation result is output.

[0021] Preferably, the initial evaluation results are fed back to the judge's interactive interface to obtain the judge's weight modification instructions for each dimension, including:

[0022] Displaying the initial assessment results in a visual form on an interactive interface, highlighting the contribution ratio of legal applicability and social impact;

[0023] receiving an adjustment signal input by a judge, wherein the adjustment signal reflects a preference for legal application or social impact weight;

[0024] A weight offset is generated according to the adjustment signal, and a normalization constraint is performed so that the sum of the adjusted weights remains 1.

[0025] Preferably, according to the weight correction instruction, the calculation parameters of the legal application dimension and the social influence dimension are dynamically adjusted, including:

[0026] The offset direction and amplitude in the weight correction instruction are analyzed to determine the target allocation ratio of the legal application dimension and the social influence dimension;

[0027] According to the target allocation ratio, the internal parameters for dimension fusion are updated, so that the new parameter value meets the current weight requirement;

[0028] The discretion interval boundary of the legal application dimension is subjected to weighted scaling, and the scaling ratio is determined by the response sensitivity factor and the adjusted weight difference;

[0029] The action intensity coefficient of the social influence dimension is synchronously updated to keep coordination with the change of the legal application dimension.

[0030] Preferably, the adjusted dimension combination is recalculated to generate the corrected evaluation result, including:

[0031] The updated discretion center value is determined based on the dynamically adjusted legal application dimension and social influence dimension;

[0032] A new confidence range is determined according to the discretion center value, and the confidence range is determined by the fluctuation amplitude and the judge preference direction;

[0033] In the confidence range, the biased correction value is calculated in combination with the judge preference direction, and the corrected evaluation result is finally output as an updated decision reference.

[0034] Preferably, the corrected evaluation result is compared with historical judgment data to form a difference analysis report, including:

[0035] A set of historical cases matching the type of the current case is extracted from the judgment database, and the set contains the final judgment value of each case;

[0036] The deviation between the corrected evaluation result and the judgment value of each historical case is calculated, and the overall deviation mean is counted;

[0037] The difference level is divided according to the deviation mean, and the difference level consists of three types: low difference, medium difference and high difference;

[0038] According to the difference level, a structured analysis report is generated, listing the key deviating cases and the recommended review direction.

[0039] Preferably, the method further includes generating an adjustment strategy based on the difference analysis report and updating the parameters in the collaborative analysis process to output the final evaluation conclusion.

[0040] On the other hand, the present invention proposes a judicial multi-dimensional evaluation and decision-making system based on deep learning, comprising:

[0041] A case feature extraction module is used to receive case elements input by the judge, generate a feature vector of the case, and extract the legal application dimension and social impact dimension of the case based on the feature vector;

[0042] An initial assessment generation module is used to generate initial assessment results based on a collaborative analysis of the legal applicability dimension and the social impact dimension. It also uses deep learning technology to analyze historical judicial data to help judges quickly find relevant precedents. The module then feeds the initial assessment results back to the judges' interactive interface, obtaining the judges' instructions on weighting adjustments for each dimension. Furthermore, the module combines natural language processing and speech recognition technology to provide support for case inquiries, information retrieval, and legal consultation through a digital human cartoon-like intelligent assistant.

[0043] A parameter dynamic adjustment module, configured to dynamically adjust the calculation parameters of the legal applicability dimension and the social impact dimension according to the weight modification instruction;

[0044] The result optimization and analysis module is used to recalculate the adjusted dimension combination, generate a revised evaluation result, compare the revised evaluation result with the historical judgment data, and form a difference analysis report.

[0045] Technical effects and advantages of the present invention: The present invention proposes a judicial multi-dimensional assessment and decision-making method and system based on deep learning, which has the following advantages over the existing technology:

[0046] This invention constructs a closed-loop evaluation process by introducing features such as feature vector generation, collaborative analysis of the legal application dimension and the social impact dimension, response to judge weight correction instructions, and differential analysis based on historical data. This method can integrate the professional judgment of judges in real time during the evaluation process and dynamically adjust model parameters to make the final output more closely aligned with actual judicial practice. This improves the adaptability and interactivity of the judicial assistance system, enabling the evaluation results to reflect the balance between the application of legal provisions and social impact, and enhancing the practicality and credibility of artificial intelligence in judicial decision-making support. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flowchart of the judicial multi-dimensional evaluation and decision-making method based on deep learning of the present invention;

[0048] Figure 2 This is a block diagram of the judicial multi-dimensional evaluation and decision-making system based on deep learning of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0050] The present invention provides Figure 1 A judicial multi-dimensional evaluation and decision-making method based on deep learning is shown, comprising the following steps:

[0051] Step 1: Receive case elements input by the judge and generate a feature vector for the case; including the following steps:

[0052] Natural language processing (NLP) technology is used to semantically analyze the case elements input by the judge to extract the types of behavior and subject relationships involved in the case. Specifically, word segmentation and part-of-speech tagging technology is first used to identify keywords, and then pre-trained legal field semantic models (such as Legal-BERT or LSTM-based classification networks) are combined to identify and extract the action verbs, subjects involved in the case and their relationships in the text.

[0053] Establish a unified case behavior coding system, for example, using a hierarchical coding method (such as a tree structure), with each behavior type corresponding to a unique numerical identifier (such as "contract dispute = 101", "personal injury = 203"). Based on the behavior types extracted in the previous step, convert them into corresponding numerical codes and arrange them into vector form according to fixed dimensions. Map the behavior types to numerical identifiers in the preset coding system to form the basic description vector A;

[0054] Determine the responsibility weights between the parties based on the subject relationship and calculate the associated impact factors Where p1 is the degree of positional deviation expressed by the plaintiff in their statement, typically ranging from [0 to 1], with larger values ​​indicating a stronger position; p2 is the degree of positional deviation expressed by the defendant in their statement, also ranging from [0 to 1]. ci reflects the degree of conflict or coordination between the parties involved in the case, with higher values ​​indicating more opposing positions and a more significant impact. By measuring the degree of positional deviation between the parties, we can capture the intensity of the conflict between the parties involved in the case, helping to introduce dynamic factors in the allocation of responsibility into subsequent models and improving the accuracy of sentencing recommendations.

[0055] Combine the basic description vector A and the associated influence factor ci, perform a weighted fusion operation, and obtain the final feature vector , where a is the weight coefficient of the basic description vector A, which controls the impact of the case behavior type on the overall characteristics; b is the weight coefficient of the associated influence factor ci, which reflects the regulatory effect of the case subject's position on the characteristic vector.

[0056] Step 2: Extracting the legal application dimension and social impact dimension of the case based on the feature vector; including the following steps:

[0057] The feature vector V is input into the preset classification framework (a pre-trained text classification model) to identify the categories of legal provisions related to the case and form a preliminary set of legal labels. This process leverages the advantages of deep learning in semantic understanding and classification tasks to improve the accuracy and efficiency of legal provision matching.

[0058] Based on a preliminary set of legal tags, the system searches for the applicable conditions and discretionary scope of corresponding provisions to construct the legal application dimension. Specifically, after obtaining the preliminary set of legal tags, the system further searches the legal knowledge base to obtain the legal provisions corresponding to each tag. The system pays special attention to extracting applicable conditions (such as the nature of the behavior and the identity of the subject) and discretionary scope (such as the range of prison terms and the amount of compensation). This information is integrated and encoded into quantifiable numerical indicators, forming the foundation of the legal application dimension.

[0059] The social impact dimension is extracted by matching related records in the social evaluation database with the case's field and legal application dimensions. Specifically, after determining the legal application dimension, the system conducts an in-depth analysis of the field involved in the case (e.g., criminal case, civil dispute, etc.). Furthermore, the system combines information such as case type, party identities, and regional background to search the social evaluation database for historical cases with similar characteristics and their accompanying social feedback data (e.g., public opinion trends, media coverage, and difficulty in enforcing judgments). This data is categorized and organized into social impact dimensions, which are used to assess the potential social response to the case.

[0060] Normalization is performed on the legal application dimension and the social impact dimension to make them comparable in the evaluation space. The processing method is as follows:

[0061] For any dimension value x, calculate its normalized result ,in Represents the historical minimum and maximum values ​​of the dimension. The principle of this formula is to linearly transform the original data and map it to the interval [0,1], thereby allowing the comparison and fusion of data of different dimensions on a unified scale. are constants obtained by analyzing historical data, which ensure the stability and consistency of the normalization process.

[0062] Step 3: Generate an initial assessment result based on a collaborative analysis of the legal application dimension and the social impact dimension. Simultaneously, use deep learning technology to analyze historical judicial data to help judges quickly find relevant precedents. This includes the following steps:

[0063] After obtaining the legal application dimension, the system automatically derives a reasonable discretionary range based on the applicable conditions of the legal provisions and case data covered by the dimension and the specific circumstances of the case. The discretionary benchmark range of the case is determined based on the legal application dimension, which is denoted as ; This interval represents the possible range of results when the current case is executed in full accordance with the legal provisions, where Indicates the lower limit value, This process relies on statistical modeling of historical judgment data and combines it with normative descriptions in the expert knowledge base to ensure the rationality and authority of the discretionary range.

[0064] According to the social impact dimension, the discretionary benchmark interval is offset and corrected to obtain the adjusted central value , where S represents the intensity of social influence and α is the empirical adjustment factor. This formula linearly shifts the midpoint of the original discretionary interval, with the offset determined by both the intensity of social influence S and the empirical adjustment factor α. It successfully integrates legal logic with social feedback, ensuring that the assessment results adhere to legal norms while adapting to the social context behind the case, thereby enhancing the flexibility and practical adaptability of decision-making support.

[0065] A floating range is set near the adjusted center value, and a confidence region is generated based on the distribution density of historical cases. This region reflects the credible range of the evaluation results, which not only retains the stability of machine calculations but also leaves room for manual intervention by judges.

[0066] Map the confidence region to an interpretable scoring system and output the initial evaluation results. The scoring is calculated as follows: ,in is the reference standard value, is a scaling factor used to control the impact of deviations on the score, keeping the score within a reasonable range. This formula converts the discretionary results into an easily understandable percentage score, allowing judges to quickly identify whether the assessment results are close to conventional precedent or recommended standards.

[0067] Step 4: Feedback the initial assessment results to the judge's interactive interface to obtain the judge's instructions on weight modification for each dimension. Combined with natural language processing and speech recognition technologies, the digital human cartoon image intelligent assistant provides support for case inquiry, information retrieval, and legal consultation, including the following steps:

[0068] The initial assessment results are displayed in a visual form on the interactive interface through a graphical interface, highlighting the contribution ratio of legal applicability and social impact; for example, a pie chart or bar chart is used to dynamically reflect the weight distribution of the two in the current assessment, helping judges quickly understand the composition logic of the model output.

[0069] Based on the visual display, the system provides interactive controls (such as sliders and plus / minus buttons) that allow judges to adjust the relative importance of the two dimensions of legal application and social impact based on their actual judgment. When judges operate these controls, the system receives the judge's input adjustment signal, which reflects the judge's preference for the weight of legal application or social impact.

[0070] The weight offset Δw is generated according to the adjustment signal, and the normalization constraint is performed so that the sum of the adjusted weights remains 1. The specific relationship is: ;in, is the original weight; is the adjusted weight; this formula adjusts the original weight through linear transformation, keeping the original distribution ratio unchanged, and introduces an intervention factor Δw to ensure that the new weight meets the normalization requirements and reflects the judge's subjective preference.

[0071] Obtaining a new weight combination The system then automatically updated the calculation rules for integrating the legal application and social impact dimensions. For example, in subsequent recalculations of the discretionary center value and revisions to assessment results, the updated weights were used in weighted averaging operations, ensuring that the entire assessment process remained synchronized with the judge's latest judgment.

[0072] Step 5: Dynamically adjust the calculation parameters of the legal application dimension and the social impact dimension according to the weight modification instruction; including the following steps:

[0073] After receiving the weight correction instruction submitted by the judge through the interactive interface, the system parses the offset direction and amplitude in the weight correction instruction, determines the target allocation ratio of the legal application dimension and the social impact dimension, and uses it to guide the subsequent parameter adjustment process.

[0074] According to the target allocation ratio, the internal parameters used for dimension fusion are updated. The adjustment method is: the original parameters are redistributed proportionally so that the new parameter values ​​meet the current weight requirements; the discretionary interval boundaries of the legal application dimension are weighted and scaled. The scaling formula is: , where γ is the response sensitivity factor, is the adjusted weight of applicable law; is the original weight of legal application; when the judge increases the weight of the legal application dimension, the system simultaneously expands the scope of the discretionary interval, making it more inclined to the rigid constraints of legal provisions; otherwise, the scope is narrowed, reflecting stronger social adaptability.

[0075] The social impact dimension's influence coefficient will be adjusted synchronously to ensure it aligns with changes in the legal application dimension. For example, when the legal application weight increases, the social impact dimension's influence coefficient may decrease accordingly to prevent any one dimension from overly dominating the entire assessment process, thereby maintaining the stability and rationality of the overall assessment.

[0076] Step 6: Recalculate the adjusted dimension combination to generate a revised evaluation result; including the following steps:

[0077] After completing the parameter adjustment of the legal application dimension and the social impact dimension, the system determines the updated discretion center value based on the dynamically adjusted legal application dimension and social impact dimension. , where S' represents the updated social impact intensity. This central value not only reflects the penalty benchmark stipulated in the legal provisions but also incorporates the latest updated social impact intensity S', controlling its influence through an empirical adjustment factor α. This formula combines legal norms with social feedback to form a more adaptable evaluation basis.

[0078] A new confidence range is drawn based on the discretionary center value, and the upper confidence limit is , the lower confidence limit is , where δ represents the amplitude of fluctuation; it can be set based on the standard deviation of historical precedents or expert experience. This method draws on the concept of confidence intervals in statistics to construct an interpretable numerical interval to reflect the uncertainty of the assessment results and provide room for subsequent manual intervention.

[0079] Calculate the bias correction value by combining the judge's preference direction within the confidence range ; When a judge is more inclined towards a certain dimension, the system should make a moderate shift in that direction within the confidence interval so that the evaluation results better match the judge's judgment style.

[0080] The final output is the revised evaluation result, R = M' + E, which serves as a reference for updated decision-making. This formula fine-tunes the originally centered discretionary value based on the judge's subjective inclination, bringing the final result closer to the user's judgment while remaining within the confidence range, ensuring the rationality and controllability of the result.

[0081] Step 7: Compare the revised assessment results with the historical judgment data to form a difference analysis report; including the following steps:

[0082] Extract a set of historical precedents that match the current case type from the judgment database, which contains the final judgment value of each case ; Calculate the deviation between the corrected evaluation result R and the judgment value of each historical case , and calculate the overall deviation mean This method draws on the idea of ​​mean absolute deviation in error analysis. By comparing the evaluation results with historical data, the degree of deviation is quantified to determine whether the model output conforms to the conventional case trend.

[0083] Divide the difference level according to the deviation mean. , it is marked as low difference, if , marked as medium difference, otherwise marked as high difference; , is the difference level threshold; this step provides the system with a qualitative description of the degree of difference, which helps judges quickly understand the level of consistency between the evaluation results and historical judgments, and decide whether further review or adjustment is needed.

[0084] A structured analysis report is generated based on the discrepancy level, listing key deviation cases and recommended review directions to assist judges in making further decisions. The analysis report includes, but is not limited to, the overall consistency of the current assessment results with historical judgments, a list of cases with significant deviations and their specific values, an analysis of the possible causes of the discrepancy, and recommended review directions for different discrepancy levels. For example, in cases with high discrepancies, the system may prompt judges to focus on whether the social impact weighting is reasonable or whether the boundaries of legal application have shifted significantly.

[0085] Step 8: Based on the difference analysis report, generate an adjustment strategy and update the parameters in the collaborative analysis process to output the final evaluation conclusion; including the following steps:

[0086] After completing the discrepancy analysis, the system conducts a consistency check on the output logic of the current evaluation model based on the discrepancy level (low, medium, or high). This judgment mechanism comprehensively assesses whether the current evaluation results are consistent with the basic trends in judicial practice by integrating various information such as case type, decision complexity, and the intensity of judge feedback. A high discrepancy level indicates that the evaluation results deviate significantly from conventional precedent, necessitating further adjustments to the model parameters to enhance its applicability.

[0087] If the consistency level fails to meet the established standards, a parameter correction strategy needs to be formulated to adjust the integration weight of the legal application dimension and the social impact dimension. This correction strategy is dynamically adjusted based on the difference level and the direction of judges’ feedback. and , thereby making the evaluation model closer to conventional judicial logic or the trial style of a specific case.

[0088] Recalculate the discretionary center value according to the new weight ,in represents the legal adaptation score, represents the social impact score; is the final adjusted weight; this formula comprehensively considers the influence of legal norms and social factors by weighted averaging the two core evaluation dimensions to ensure that the final discretionary result meets the normative requirements while maintaining flexibility.

[0089] After executing all optimization steps, the system will output the final discretionary center value The conclusion is presented as a numerical value and is accompanied by a concise explanatory document covering the evaluation path, changes in key parameters, and comparisons with previous precedents, so that the judge can fully grasp the information and make the final decision accordingly.

[0090] On the other hand, the present invention proposes a judicial multi-dimensional evaluation decision system based on deep learning, such as Figure 2 Shown, including:

[0091] A case feature extraction module is used to receive case elements input by the judge, generate a feature vector of the case, and extract the legal application dimension and social impact dimension of the case based on the feature vector;

[0092] An initial assessment generation module is used to generate initial assessment results based on a collaborative analysis of the legal applicability dimension and the social impact dimension. It also uses deep learning technology to analyze historical judicial data to help judges quickly find relevant precedents. The module then feeds the initial assessment results back to the judges' interactive interface, obtaining the judges' instructions on weighting adjustments for each dimension. Furthermore, the module combines natural language processing and speech recognition technology to provide support for case inquiries, information retrieval, and legal consultation through a digital human cartoon-like intelligent assistant.

[0093] A parameter dynamic adjustment module, configured to dynamically adjust the calculation parameters of the legal applicability dimension and the social impact dimension according to the weight modification instruction;

[0094] The result optimization and analysis module is used to recalculate the adjusted dimension combination, generate a revised evaluation result, compare the revised evaluation result with the historical judgment data, and form a difference analysis report.

[0095] In addition, the above modules are also used to implement other steps of the above-mentioned judicial multi-dimensional evaluation and decision-making method based on deep learning during execution, as shown in the following embodiments:

[0096] A court was hearing a contract dispute case in which the plaintiff (a property management company) sued the defendant (a real estate company) for 2.77 million yuan in property management service fees and 1.21 million yuan in liquidated damages. The judge hoped to use this system to assist in generating a ruling recommendation.

[0097] Step 1: Generate case feature vector

[0098] Semantic parsing and encoding: After the judge enters the case description, the system uses NLP technology to extract keywords such as "contract dispute," "property service fee," and "liquidated damages." The Legal-BERT model identifies the behavior type as "contract dispute = 101" and the subject relationship as "property company - enterprise = 302." The plaintiff's position deviation (p1) is 0.8 (strongly advocating for liquidated damages), while the defendant's position deviation (p2) is 0.5 (partially admitting liability).

[0099] Correlation impact factor calculation: Correlation impact factor =(0.8+0.5) / 2= 0.65.

[0100] Feature vector generation: Basic description vector A = [101, 302] (encoded behavior type and subject relationship). Weight coefficient a = 0.7 (focus on behavior type), b = 0.3 (focus on stance influence). Final feature vector =0.7*[101,302]+0.3*0.65= [70.7,211.4,0.195].

[0101] Step 2: Extracting the legal applicability and social impact dimensions

[0102] Legal application dimension: Matching legal provisions through classification models. Discretion range =[2 million, 3.5 million] (refer to historical precedents).

[0103] Social impact dimension: The social impact intensity of similar cases in the social evaluation database is S=0.7 (high public attention).

[0104] Normalization: Normalization of the legal application dimension: z1 = (200-100) / (350-100) = 0.4. Normalization of the social impact dimension: z2 = (0.7-0) / 1 = 0.7.

[0105] Step 3: Generate initial evaluation results

[0106] Discretion center value calculation:

[0107] =(200+350) / 2+0.50*7=275+0.35= 2.7535 million yuan (α=0.5).

[0108] Confidence Region and Scoring:

[0109] Fluctuation range δ = 100,000 yuan, confidence interval =[285.35,265.35].

[0110] Reference standard value ref=2.7 million yuan;

[0111] score =100-1*5.35= 94.65 points (β=1).

[0112] Step 4: Judge interaction and weight adjustment

[0113] Visual display: The interface shows that legal applicability contributes 60% and social impact contributes 40% (initial weights w1=0.6, w2=0.4).

[0114] Judge adjusts the signal: The judge believes that social influence should be more prominent, and the slider is adjusted to Δw=0.1 (increasing the weight of social influence).

[0115] Weight update: w1'=0.6-0.1= 0.5, w2'=0.4+0.1= 0.5.

[0116] Step 5: Dynamically adjust parameters

[0117] Discretion interval scaling: γ = 0.2 (response sensitivity factor), =200*(1+0.2*(0.5-0.6))=200*(1-0.02)= 1.96 million yuan.

[0118] Adjustment of social impact intensity: After the social impact weight is increased, S'=0.7*1.1=0.77 (the coefficient is increased by 10% simultaneously).

[0119] Step 6: Correct the evaluation results

[0120] Updated discretionary center value: M'=(196+350) / 2+0.5*0.77=273+0.385= 2.73385 million yuan.

[0121] Bias correction: Bias correction value E = (283.385 - 263.385) * (0.5 - 0.5) / 2 = 0 (no bias when weights are equal). Corrected result R = 273.385 + 0 = 2,733,850 yuan.

[0122] Step 7: Variance Analysis Report

[0123] Comparison of historical cases: Extract 5 similar cases and the judgment value They are 2.6 million, 2.75 million, 2.8 million, 2.9 million and 3 million respectively.

[0124] deviation They are 13.385, -1.615, 6.615, 16.615 and 26.615 respectively.

[0125] Overall deviation mean =(13.385+1.615+6.615+16.615+26.615) / 5= 130,900 yuan.

[0126] Differentiation level classification: setting =100,000 yuan, =200,000 yuan→ =13.09∈ medium difference.

[0127] Analysis report: Mark cases with large deviations (such as the 3 million judgment) and recommend reviewing whether the social impact weight is too high.

[0128] Step 8: Generate final evaluation conclusion

[0129] Parameter correction: Adjust weights based on medium variance: (Applicable Law), =0.45 (social influence).

[0130] Final discretionary center value:

[0131] =0.5 (legal adaptation score), =0.8 (social impact score).

[0132] =0.550.5+0.450.8=0.275+0.36= 0.635 (normalized value).

[0133] Output conclusion:

[0134] Will Denormalized to actual amount: 0.635*(350-200)+200= 2.9525 million yuan.

[0135] Final judgment recommendation: The system outputs a final evaluation conclusion of 2.9525 million yuan. The judge, combined with the difference analysis report, decides to adopt the recommendation and adjust the preservation measures, unfreeze the defendant's bank account, seal and replace the real estate, taking into account both legal norms and the survival needs of the enterprise.

[0136] This case demonstrates the entire process from case input to judgment recommendation. By dynamically adjusting weights and parameters, the system achieves a balance between legal logic and social feedback, improving the efficiency and rationality of judicial decision-making.

[0137] To enhance the user experience, the system also integrates voice recognition and speech synthesis capabilities, allowing users to interact with the AI ​​assistant via voice. Users can start, continue, or restart a conversation, and the AI ​​assistant will respond in real time based on the user's input and stream it back, ensuring the timeliness and accuracy of information delivery. This approach significantly improves the efficiency and fairness of judicial evaluation decisions, providing strong support for modern judicial practice.

[0138] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A judicial multi-dimensional evaluation and decision-making method based on deep learning, characterized by: The following steps are involved: Receive case elements input by the judge, generate a feature vector of the case, and extract the legal application dimension and social impact dimension of the case based on the feature vector; After collaboratively analyzing the legal applicability and social impact dimensions, an initial assessment result is generated. Deep learning technology is used to analyze historical judicial data to help judges quickly find relevant precedents. The initial assessment result is fed back to the judge's interactive interface to obtain the judge's instructions on how to adjust the weights of each dimension. Combined with natural language processing and speech recognition technology, a digital human cartoon avatar intelligent assistant provides support for case inquiries, information retrieval, and legal consultation. According to the weight modification instructions, the calculation parameters of the legal application dimension and the social impact dimension are dynamically adjusted; Recalculating the adjusted dimension combination to generate a revised evaluation result, comparing the revised evaluation result with the historical judgment data, and forming a difference analysis report; The step of generating the feature vector of the case includes: Perform semantic analysis on the case elements input by the judge to extract the behavior types and subject relationships involved in the case; Mapping the behavior type to a numerical identifier in a preset coding system to form a basic description vector; Determine the responsibility weights between the parties based on the subject relationship, calculate the correlation influence factor, combine the basic description vector and the correlation influence factor, and generate a final feature vector through a weighted fusion operation; After conducting a collaborative analysis based on the legal applicability and social impact dimensions, the initial assessment results are generated, including: Determine a case discretion benchmark interval based on the legal application dimension, and perform offset correction on the discretion benchmark interval based on the social impact dimension to obtain an adjusted central value; A floating range is set around the adjusted central value, and a confidence region is generated based on the historical case distribution density. The confidence region is mapped into an interpretable scoring system, and the initial evaluation result is output.

2. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: The legal application dimension and social impact dimension of the case are extracted based on the feature vector, including: Inputting the feature vector into a preset classification framework to identify the categories of legal provisions related to the case and form a preliminary legal label set; Based on the preliminary legal label set, find the applicable conditions and discretionary scope of the corresponding provisions and construct the legal application dimension; Based on the field to which the case belongs and the legal application dimensions, the related records in the social evaluation database are matched and the social impact dimensions are extracted.

3. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: Feedback the initial evaluation results to the judge's interactive interface to obtain the judge's weight modification instructions for each dimension, including: Displaying the initial assessment results in a visual form on an interactive interface, highlighting the contribution ratio of legal applicability and social impact; receiving an adjustment signal input by a judge, wherein the adjustment signal reflects a preference for legal application or social impact weight; A weight offset is generated according to the adjustment signal, and a normalization constraint is performed so that the sum of the adjusted weights remains 1.

4. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: According to the weight modification instructions, the calculation parameters of the legal application dimension and the social impact dimension are dynamically adjusted, including: Analyze the deviation direction and magnitude in the weight modification instruction to determine the target allocation ratio between the legal application dimension and the social impact dimension; According to the target allocation ratio, the internal parameters for dimension fusion are updated so that the new parameter values ​​meet the current weight requirements; Perform weighted scaling on the boundaries of the discretionary interval of the legal application dimension, where the scaling ratio is determined by the difference between the response sensitivity factor and the adjusted weight; The effect intensity coefficient of the social impact dimension is updated synchronously to keep it coordinated with the changes in the legal application dimension.

5. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: Recalculate the adjusted dimension combination to generate a revised assessment result, including: Determine the updated discretion center value based on the dynamically adjusted legal application dimension and social impact dimension; A new confidence range is drawn based on the central value of the discretion, and the confidence range is determined by the fluctuation range and the direction of the judge's preference; The bias correction value is calculated based on the judge's preference direction within the confidence range, and the corrected evaluation result is finally output as a reference for updated decision-making.

6. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: Compare the revised assessment results with historical judgment data to form a difference analysis report, including: Extracting a historical case set matching the current case type from a judgment database, wherein the set includes a final judgment value of each case; Calculate the deviation between the revised evaluation result and the judgment value of each historical case, and calculate the overall mean deviation; Divide the difference level according to the deviation mean, and the difference level consists of three categories: low difference, medium difference and high difference; Generate a structured analysis report based on the difference level, listing key deviation cases and recommended review directions.

7. The judicial multi-dimensional evaluation and decision-making method based on deep learning according to claim 1 is characterized in that: The method further comprises: Based on the difference analysis report, generate adjustment strategies and update parameters in the collaborative analysis process, and output the final evaluation conclusion; The decision-making support system developed based on deep learning and big data technology provides judges with detailed case analysis, specific sentencing suggestions, and legal application recommendations to support the final ruling; Throughout the entire judicial service process, including case acceptance, court trial, and judgment execution, AI technology is used to provide full-process intelligent services, enhance judicial transparency and public trust, and ensure that each stage is carried out efficiently and fairly.

8. A judicial multi-dimensional evaluation and decision-making system based on deep learning for implementing the method according to any one of claims 1 to 7, characterized in that: include: A case feature extraction module is configured to receive case elements input by a judge, generate a feature vector for the case, and extract the legal applicability dimension and social impact dimension of the case based on the feature vector. Generating the feature vector for the case includes: semantically parsing the case elements input by the judge to extract the behavior types and subject relationships involved in the case; mapping the behavior types to numerical identifiers in a preset coding system to form a basic description vector; determining the responsibility weights between the parties based on the subject relationships, calculating the associated impact factors, and combining the basic description vector with the associated impact factors to generate a final feature vector through a weighted fusion operation. An initial assessment generation module is used to generate an initial assessment result after collaborative analysis based on the legal application dimension and the social impact dimension, and at the same time use deep learning technology to analyze historical judicial data to help judges quickly find relevant precedents, and feed back the initial assessment result to the judge's interactive interface to obtain the judge's weight correction instructions for each dimension, and combine natural language processing and voice recognition technology to provide case query, information retrieval, and legal consultation support through a digital human cartoon image intelligent assistant; wherein, the initial assessment result is generated after collaborative analysis based on the legal application dimension and the social impact dimension, including: determining the discretionary benchmark interval of the case based on the legal application dimension, offset-correcting the discretionary benchmark interval according to the social impact dimension, and obtaining the adjusted central value; setting a floating range near the adjusted central value, generating a confidence region based on the distribution density of historical precedents, mapping the confidence region into an interpretable scoring system, and outputting the initial assessment result; A parameter dynamic adjustment module, configured to dynamically adjust the calculation parameters of the legal applicability dimension and the social impact dimension according to the weight modification instruction; The result optimization and analysis module is used to recalculate the adjusted dimension combination, generate a revised evaluation result, compare the revised evaluation result with the historical judgment data, and form a difference analysis report.

Images