Data-driven criminal period prediction system

Abstract

The invention relates to a data-driven criminal period prediction system, belongs to the technical field of artificial intelligence, solves the problems of insufficient accuracy, poor regional adaptability and the like of prisoner classification and criminal period prediction in the prior art, and provides a multi-model fusion provincial prisoner classification and criminal period prediction system based on real judicial sentencing logic. The system comprises an input display module, a legal judgment element extraction module, a sentency and insignment prediction module, a sentency judgment prediction module and a judgment output display module, and can effectively improve the prediction precision and optimize judicial decision support by fusing multiple advanced algorithms and combining judicial practices of different provinces. The method has a wide application prospect and a strong regional adaptation capability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a data-driven sentence prediction system. This system combines a large legal model with regular expression matching technology to extract legal details from court documents. The quantified features are then input into a multi-model integrated sentencing system, ultimately outputting a classification of criminal cases and a predicted sentence. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent applications in the legal field have gradually become a research hotspot. In judicial practice, high professional requirements and a heavy workload in case handling have become real factors accelerating the development of intelligent judicial processes. Meanwhile, the vast amount of publicly available court judgments provides both professional and substantial research data for intelligent judicial processes, containing a wealth of information worth exploring. Therefore, research has emerged on using big data and machine learning technologies to assist judges in sentencing decisions based on court judgments and other legal documents. This research not only has the potential to improve judicial efficiency but also to promote fairness and transparency within the legal system. However, the inherent complexity of legal language and the diversity of case details present significant challenges in handling nuanced tasks such as sentencing prediction.

[0003] Currently, automated analysis of legal judgments and sentence prediction systems primarily rely on traditional rule matching and simple regression models. These methods cannot effectively capture the nonlinear relationships within legal judgments, thus limiting their ability to handle complex legal situations. Furthermore, there is a lack of research on the crucial aspect of sentence type prediction, and no studies comprehensively consider the relationship between probation and fixed-term imprisonment. In addition, existing systems fail to account for legal differences across provinces, ignoring regional variations in legal practice, resulting in low prediction accuracy.

[0004] For example, Chinese patent CN109241528B, entitled "A Method, Device, Equipment, and Storage Medium for Predicting Sentencing Outcomes," primarily provides a method for automatically predicting sentencing outcomes based on non-judgment documents of a specified case using a pre-established model. By acquiring non-judgment documents, extracting sentencing elements, and inputting them into the prediction model, it can assist judges in making decisions. Although this method can automatically predict sentencing outcomes, its accuracy may be affected by incomplete information in non-judgment documents or limitations in model training data. Furthermore, it does not fully consider the impact of differences in legal practices across different regions on sentencing.

[0005] Chinese patent CN112163707B, entitled "A Method for Predicting Sentences Based on Bayesian Networks," primarily predicts sentences by preprocessing data from tried court documents, learning the structure of a Bayesian network, and learning its parameters to construct a sentence prediction model. This method can identify and explain key factors affecting sentence length, overcoming the subjectivity of expert-constructed networks. However, the network construction process involves multiple complex algorithms, requiring significant data and computational resources, and the accuracy of predictions may be affected by data bias.

[0006] Chinese patent CN114239939B, entitled "A Method for Predicting Sentences by Integrating Auxiliary Knowledge," primarily achieves sentence prediction through steps such as preprocessing legal document datasets, integrating auxiliary knowledge, and building a hybrid network model. This method addresses the problem of uneven dataset distribution, improving the convenience, accuracy, and efficiency of prediction. However, the model relies on specific datasets and complex processing procedures; its generalization ability across different datasets may need further verification, and its adaptability to newly emerging legal situations may be insufficient.

[0007] Chinese patent CN114860900B, entitled "A Sentencing Prediction Method and Device," primarily involves acquiring case-related information and textual descriptions of criminal facts, performing vectorization processing and feature extraction, and then predicting legal provisions, charges, and sentences. While this method comprehensively considers various information sources to improve prediction accuracy, it may face high computational complexity when processing long texts, and the prediction results lack interpretability. Summary of the Invention

[0008] In view of the above problems, this invention provides a data-driven sentence prediction system that solves the problems of insufficient accuracy in sentence classification and sentence prediction, and poor regional adaptability in existing technologies. It provides a multi-model fusion system and method for provincial-level sentence classification and sentence prediction based on real judicial sentencing logic. By integrating multiple advanced algorithms and combining them with judicial practices in different provinces, this system can effectively improve prediction accuracy and optimize judicial decision support, and has broad application prospects and strong regional adaptability.

[0009] This invention provides a data-driven sentence prediction system, including an input display module, a legal judgment element extraction module, a fixed-term imprisonment prediction module, a probation sentence prediction module, and a judgment output display module;

[0010] The input display module is used to input and display judgment documents and case location information;

[0011] The legal judgment element extraction module includes a document structure splitting unit and a large-model-enhanced legal element identification unit. The document structure splitting unit divides the judgment document into the facts ascertained during the trial and the court's opinion. The large-model-enhanced legal element identification unit performs legal provision matching, keyword matching, and large language model extraction based on the key case factors and sentencing factors in the facts ascertained during the trial and the court's opinion to obtain judgment details related to sentencing and case type.

[0012] Among them, the legal provision matching method determines whether the defendant has the circumstances stipulated in the legal provision based on the legal provisions cited in the judgment document;

[0013] Keyword matching uses positive-side matching, while regular expression matching is used to extract structured and meaningful plot features.

[0014] Large language models are used to extract and determine plot features;

[0015] The fixed-term imprisonment prediction module is based on the judgment characteristics related to sentencing and case type output by the legal element identification unit enhanced by the large model, combined with the comprehensive assessment of the defendant's degree of social harm given based on the substantive requirements of probation, to obtain the predicted sentence value.

[0016] The probation term prediction module is constructed based on the preconditions for the application of probation and the substantive conditions of probation, and is used to determine the predicted probation value.

[0017] The judgment output display module generates the duration of probation and fixed-term imprisonment based on the prediction results and displays the prediction results; the prediction results include the predicted sentence and the predicted probation.

[0018] Optionally, the characteristics of the circumstances to be determined include the elements constituting the crime, the circumstances under which sentencing should be applied first, the statutory circumstances for mitigation of punishment, the circumstances for leniency, and the circumstances for severity of punishment.

[0019] Optionally, the substantive conditions for probation include the severity of the crime, remorse, risk of recidivism, and impact on the community.

[0020] Optionally, the probation period prediction module includes a probation determination unit and a probation period unit.

[0021] Optionally, it also includes a regional difference processing module, which adjusts the parameter vectors in the fixed-term imprisonment term prediction module and the probation term prediction module according to the differences in legal practices in different provinces, so that the model can adapt to the legal environment of different regions.

[0022] Optionally, the expression for the fixed-term imprisonment sentence prediction module is:

[0023]

[0024] in, This represents the length of the sentence for the k-th case; This represents the saturation function of the fixed-term imprisonment prediction module for the k-th case; This serves as the starting point for sentencing in the k-th case; b represents the nth sentencing factor affecting the base sentence in the kth case, where n = 1, 2, ..., m1; n m1 represents the weighting parameter corresponding to the nth sentencing factor; m2 represents the total number of sentencing factors affecting the base sentence; m3 represents the number of prior sentencing features stipulated in the law; and m4 represents the number of other unconsidered features. p represents the j-th prior sentencing characteristic stipulated in the sentencing guidelines in the k-th case; j For the unknown weight parameters corresponding to the j-th prior applicable sentencing feature; q represents the t-th unconsidered feature in the k-th case; t Let be the weight parameter for the t-th unconsidered feature; This is the bias term for the k-th case; Represents a real number.

[0025] Optionally, it also includes a type of punishment determination unit, used to determine the type of punishment that the defendant should be punished based on the obtained characteristics of the circumstances, thereby obtaining a type of punishment classification.

[0026] Compared with the prior art, the present invention has at least the following beneficial effects:

[0027] (1) Efficient and accurate legal element identification: This invention, by combining large-scale models with regular expression matching technology, can efficiently extract complex legal plot features from judgment documents, especially those complex plots that require contextual semantic understanding. Compared with traditional methods, large-scale models can better capture the complex information implicit in legal documents, while regular expression matching technology can quickly and efficiently extract key information in fixed formats, such as case numbers and dates. This efficient feature extraction method greatly improves the system's ability to process judgment documents, ensuring the accuracy and comprehensiveness of predictions.

[0028] (2) Criminal type judgment unit: This invention adopts a highly reliable adaptive identification algorithm, which combines non-convex optimization method and recursive learning technology to cope with limited data. In order to deal with the problem of high imbalance of criminal type categories, a two-step adaptive classification algorithm is adopted, and its convergence theory is established under general data conditions including non-stationary and related samples. Finally, it can accurately classify criminal types such as: control, detention, fixed-term imprisonment, life imprisonment and death penalty.

[0029] (3) Probation Sentence Prediction Module: This invention closely integrates with judicial sentencing logic and proposes a nonlinear mathematical model for predicting probation terms. The model's input includes not only features affecting the actual sentence but also legal elements influencing probation, such as whether the defendant repents, meets the requirements for community correction, and has a risk of recidivism. By integrating multiple features, it accurately characterizes the key factors affecting probation during judicial discretion, making sentencing predictions more interpretable and easier for judicial personnel to understand and trust the prediction results. Unlike traditional black-box models, this invention's model clearly demonstrates the various factors affecting sentencing, improving the transparency and rationality of the prediction results and providing more transparent decision support for discretionary supervision.

[0030] (4) Fixed-term imprisonment prediction model: A saturated nonlinear sentencing model that conforms to the logic of judicial sentencing is adopted to refine and quantify each stage of the sentencing process. First, based on the determination of the starting point and base sentence according to the facts of the crime, the relevant characteristics of conviction are used to adjust the base sentence through a product, thereby reflecting the priority application of the relevant characteristics of conviction; then, based on other legal circumstances, an adjustment strategy of "adding in the same direction and subtracting in the opposite direction" is adopted to accurately reflect the positive and negative impact of different circumstances on the term of imprisonment; finally, a saturation function is introduced to ensure that the final predicted term of imprisonment strictly falls within the statutory range. The phased nonlinear sentencing model not only significantly improves the accuracy of term of imprisonment prediction, but also provides a reliable basis for the interpretation of the judgment, which has obvious advantages compared with existing models.

[0031] (5) Regional Difference Handling: This invention utilizes a regional difference handling module to dynamically adjust model parameters based on the differences in legal practices across different provinces. This module can optimize the model according to the characteristics of judicial practices in various regions, thereby improving the prediction accuracy in different areas. For example, some provinces have stricter sentencing for specific types of crimes, while others are relatively more lenient. Through this module, the model can adapt to the legal environment of different regions, ensuring the local adaptability and accuracy of the sentence prediction results. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the data-driven sentence prediction system of the present invention. Detailed Implementation

[0033] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0034] A specific embodiment of the present invention, such as Figure 1 A data-driven sentence prediction system is disclosed, including an input display module, a legal judgment element extraction module, a sentence type judgment unit, a fixed-term imprisonment prediction module, a probation judgment prediction module, a regional difference processing module, and a judgment output display module.

[0035] The input display module is used to input and display judgment documents and case location information;

[0036] The legal judgment element extraction module includes a document structure splitting unit and a large model-enhanced legal element recognition unit.

[0037] Furthermore, the geographical information of the case includes identity / municipality.

[0038] The document structure is divided into two parts: the facts ascertained during the trial and the court's opinion.

[0039] Furthermore, the investigation provides a detailed description of the facts of the case, including information on the criminal act, the defendant's motives, and the methods used. This information is a key factor in determining the base sentence and influencing the pronounced sentence. This court holds that this section represents some of the sentencing factors explicitly considered by the court during the judgment process. Moreover, the relevant legal provisions listed, such as the Criminal Law of the People's Republic of China and the Criminal Procedure Law of the People's Republic of China, clearly stipulate sentencing factors relevant to the case. These key case factors and sentencing factors are crucial in determining the defendant's sentence.

[0040] Among them, the large-scale model-enhanced legal element identification unit includes a large language model, keyword matching, and legal provision matching; the large-scale model-enhanced legal element identification unit performs legal provision matching, keyword matching, and large language model extraction based on the key case factors and sentencing factors ascertained during the trial and in the court's opinion, to obtain judgment features related to sentencing and case type.

[0041] Furthermore, the large language model is the DeepSeek-R1 large language model.

[0042] Specifically, the characteristics of the circumstances include the elements constituting the crime, the circumstances under which sentencing is applied first, the statutory mitigating circumstances, the circumstances for lighter punishment, and the circumstances for heavier punishment.

[0043] Exemplarily, the key factors of the case are the degree of injury of the victim, whether the defendant surrendered himself, and whether the defendant actively compensated and obtained understanding; those related to sentencing refer to the elements common to all cases in the criminal law, including: whether the defendant surrendered himself, whether he actively compensated, and whether he has a criminal record, etc.; those related to the type of case are the case characteristic factors for different charges. For example, the crime of causing traffic accidents includes whether the defendant took drugs, whether he drove without a license, and whether he complied with traffic regulations, etc., while the crime of intentional injury includes sentencing factors such as whether the defendant used a weapon and whether it was a family dispute.

[0044] Furthermore, the legal provision matching determines whether the defendant has the circumstances stipulated in the legal provision according to the legal provisions cited in the judgment document. For example, for the circumstance of abetting a person under the age of eighteen to commit a crime, different judges may use different descriptions when writing the judgment, which brings great challenges to the extraction work. To solve this problem, the present invention adopts the method of reverse matching, corresponding the legal provision to the corresponding circumstances. The legal provision matching method is used to identify the relevant legal provisions involved in the judgment document. If a certain legal provision appears in the text, it is determined that the case includes the judgment circumstance features corresponding to the legal provision.

[0045] Furthermore, the keyword matching adopts positive and negative matching, and the regular expression matching method is used to extract the judgment circumstance features with structured and clear meanings. According to the requirements of different features, corresponding regular expressions are designed, and the regular expressions are used to match the judgment document to extract the judgment circumstance features that meet the rules. For example, for the feature of "hit-and-run" in the crime of causing traffic accidents, first use re.search('逃逸',data) to find whether the text contains "逃逸", if found, it means that the "hit-and-run" behavior is matched. Then use re.search(r'(非|不|未|无|没)[^,。,;;]*?逃逸',data) to check whether there is a negative word before "逃逸", if the negative word is matched, it means that the "hit-and-run" is not constituted.

[0046] Furthermore, the large language model is used to extract complex judgment circumstance features that require semantic understanding, including a text preprocessing and chunking module, a Prompt design and model input module, a model inference and JSON format result generation module, and a JSON result parsing and circumstance feature extraction module. When using it, first chunk the text of "found through trial" and "held by this court": chunk the text of "found through trial" and "held by this court" by paragraph to form a text chunk set D; then, design corresponding prompts for each complex judgment circumstance feature and define a structured output template, and input the above text chunks into the locally deployed large language model in sequence to guide the generation of JSON format results; finally, parse the JSON format results to extract the final judgment circumstance features.

[0047] Specifically, the text preprocessing and segmentation module is used to divide the judgment document into a set of text blocks D = {d1, d2, ..., d...}. n}, where d i This represents the i-th text block. Each text block includes a segment of information from a court document. Long texts are broken down into smaller paragraphs or information blocks so that the large model can understand and process them step by step.

[0048] Specifically, the Prompt design and model input module is used to design corresponding prompts for each type of judgment plot feature, facilitating the large language model's recognition and extraction of complex legal element features. These prompts are designed based on preset judgment plot feature types, ensuring that the model can understand the task's context and accurately output the corresponding plot.

[0049] For example, suppose we need to identify the plot feature of "number of robberies," and design it as follows: "Please process the text strictly according to the following requirements: the rule for calculating the number of robberies is that each robbery is counted as 1, regardless of whether it is successful or not; consecutive robberies of different targets are considered as multiple robberies; accumulate the number of all explicitly described robberies. Output requirements: JSON key names must be Chinese characters for 'robbery number'; numeric types must be integers (without currency symbols / units); fill in 0 if the number is missing; comments and extra fields are prohibited. Please generate JSON strictly according to the above rules." Through a carefully designed prompt, the model is guided to accurately identify and extract the specific legal plot.

[0050] Specifically, the model inference and JSON-formatted result generation module is used to feed the text block set D and the corresponding prompts into the locally deployed large language model. The model performs inference based on the prompts and the text block set, generating JSON-formatted results, which include the extracted decision plot features.

[0051] Specifically, the JSON result parsing and plot feature extraction module is used to parse the generated JSON format results and extract specific plot features for judgment.

[0052] Traditional regular expression methods excel at efficiently processing structured or semi-structured data. Since judgment documents typically follow standardized formats, regular expression matching improves efficiency when extracting fixed-format features (such as case numbers, dates, and predefined legal provisions). However, regular expression methods encounter limitations when handling more complex tasks, such as identifying the severity of injury in legal cases. In contrast, large language models possess powerful semantic understanding capabilities and the ability to adapt to different legal texts, enabling them to extract context-dependent features. However, they can produce errors when extracting features that are difficult to describe in language, and their high computational cost and slow processing speed pose significant challenges. Therefore, this invention combines regular expressions and large language models, fully leveraging the advantages of both while mitigating their respective shortcomings, thereby achieving more accurate and efficient feature extraction from large-scale judicial documents.

[0053] For example, taking the Criminal Law of the People's Republic of China and the Sentencing Guidelines for Common Crimes as examples, the characteristics of the sentencing circumstances are divided into five key circumstances: circumstances constituting the crime, circumstances applicable to sentencing prior to the crime, statutory mitigating circumstances, circumstances for leniency, and circumstances for severity. First, by analyzing the facts of the case and the court's judgment in the judgment documents, circumstances constituting the crime are extracted to determine whether they meet the elements of a specific crime. Next, based on factors such as the defendant's age and whether they were an accomplice in the crime, circumstances applicable to sentencing prior to the crime are identified. Then, according to the specific provisions of the Criminal Law, it is identified whether statutory mitigating circumstances exist, such as surrender and meritorious service. Next, considering the specific circumstances of the case, it is determined whether circumstances for leniency apply, such as whether the defendant actively compensated the victim for their losses. Finally, based on the special severity of the crime and whether the defendant has a prior criminal record, it is determined whether circumstances for severity apply. Through this process, this invention can comprehensively and systematically extract and analyze sentencing-related characteristics to assist in sentencing decisions.

[0054] This invention utilizes a large-model-enhanced legal element identification unit to select sentencing-related features from the investigation and the court's opinion, based on the charge. These features encompass the circumstances of the defendant and the victim, as well as the defendant's criminal conduct. These attributes are factors considered by judges during the sentencing process and conform to legal sentencing logic.

[0055] Some existing technologies employ word embedding techniques to extract key features. However, this method may lead to incomplete keyword extraction or inaccurate feature recognition, especially when the description in the judgment document is vague or imprecise. For example, a judgment might state: "Regarding the defense counsel's claim that the defendant acted in self-defense, this court does not accept it." In this case, the model might correctly identify the phrase "self-defense," but fail to capture the subsequent negation, resulting in an error in feature extraction.

[0056] Furthermore, the punishment type determination unit, based on the obtained characteristics of the circumstances (such as the basic facts of the crime, the consequences of the crime, etc.), determines what type of punishment the defendant should be punished with, thereby obtaining the punishment type classification;

[0057] Furthermore, a data-driven sentence prediction system is constructed from an input display module, a legal judgment element extraction module, a sentence type judgment unit, a regional difference processing module, and a judgment output display module. This system may also include a fixed-term imprisonment prediction module and a probation judgment prediction module.

[0058] The core of this module is to construct a suitable crime prediction model and design a corresponding adaptive identification algorithm to train and update the model parameters. The adaptive identification algorithm of this invention can converge globally under general data conditions (without requiring the classic independent and identically distributed condition), making it particularly suitable for analyzing complex judicial sentencing data with interconnected contexts.

[0059] Furthermore, the expression for the type of punishment determination unit is:

[0060]

[0061] Among them, y k This indicates the type of punishment for the k-th case. Let represent the regression vector composed of the judgment features of the k-th case, and d represent the total number of judgment features. Represent real numbers; Represents random noise in the k-th case; The factor representing the degree of influence of each judgment circumstance feature on the type of punishment is the unknown parameter vector to be estimated for the type of punishment judgment unit; This represents the saturation function of the crime type classification module for the k-th case.

[0062] Non-decreasing time-varying saturation function The expression is:

[0063]

[0064] Where m represents the total number of types of punishment, including control, detention, fixed-term imprisonment, life imprisonment, and death penalty; c ik The threshold for classifying the i-th type of crime in the k-th case.

[0065] Furthermore, when training the crime type prediction model (training to obtain the unknown parameter vector θ to be estimated), the crime type classification algorithm is designed based on the following loss function J, which represents the mean absolute error between the predicted category and the true category, expressed as:

[0066]

[0067] Among them, yk This indicates the actual type of sentence handed down to the defendant in the k-th case; This represents the predicted sentence category for the k-th case, where K represents the total number of cases currently being processed.

[0068] Furthermore, when training the crime prediction model, a two-step adaptive identification algorithm is used to recursively update the influence factor θ. k The second step combines the estimated impact factor values ​​from the first step. Update the influence factor θ k This continues until global convergence. By adjusting the adaptive gain matrix, adaptive factor, and adaptive step size factor, convergence is accelerated while maintaining overall convergence. Unlike traditional methods, the projection operation in this algorithm ensures that parameter updates remain within a reasonable range, thus avoiding excessive fluctuations or instability. The two steps are as follows:

[0069] Step 1: Use a recursive method to obtain the estimated value of the influence factor. The expression is:

[0070]

[0071] in, This represents the estimated value of the influence factor for the k-th case; This represents the estimated value of the influence factor for the (k-1)th case; This represents the adaptive gain of the estimated value for the k-th case; φ represents the adaptive gain matrix for the estimated value of the k-th case; k Let represent the regression vector of the k-th case; This indicates the direction of the update of the impact factor estimate for the k-th case; This represents the projection operator for the estimated values ​​of k cases.

[0072] Furthermore, the projection operator for the k-th case The expression is:

[0073]

[0074] in, express r is a bounded closed convex set; The point in the middle that minimizes the projection operator; The adaptive gain matrix for the estimated value of the k-th case. The inverse matrix.

[0075] Furthermore, the initial impact factor estimate and positive definite adaptive gain matrix You can select any option.

[0076] Furthermore, the update direction of the impact factor estimate for the k-th case. The expression is:

[0077] Where F(.) represents the probability distribution function of normal noise.

[0078] In addition, the estimated value of the k-th case is the adaptive gain matrix. The expression is:

[0079]

[0080] in, Let represent the adaptive gain matrix for the (k-1)th case; This represents the adaptive gain step size factor for the k-th case.

[0081] Furthermore, the estimated value of the k-th case has an adaptive gain. The expression is:

[0082]

[0083] in, This represents the step size parameter for estimating the k-th case.

[0084] Furthermore, the step size parameter for the k-th case has the following range of values:

[0085]

[0086] Among them, inf k≥0 This represents the step size parameter for any case k ≥ 0. The lower bound; sup k≥0 This represents the step size parameter for any case k ≥ 0. The upper boundary.

[0087] Furthermore, the estimated value of the k-th case has an adaptive gain step size factor. The expression is assumed to be:

[0088]

[0089] Among them, G k Denotes a bounded closed set of k cases. The upper bound of the norm of the inner element is: c 1k and c mk Let f(.) represent the threshold values ​​for classifying the first type of crime in the k-th case and the threshold value for classifying the m-th type of crime, respectively; f(.) represents the conditional density function of normal noise; x represents the value of...

[0090] A point that satisfies the norm inequality in the above formula; inf represents the infimum.

[0091] Step 2: Based on the estimated value of the influence factor Obtain the degree of influence factor;

[0092] Furthermore, the expression for the degree of influence factor is:

[0093]

[0094] Where, θ k θ represents the influence factor of the k-th case; k-1 This represents the influence factor of the (k-1)th case; Denotes the projection operator for the k-th case; a k P represents the adaptive gain for k cases; k Let the adaptive gain matrix of the k-th case be represented. P is the adaptive gain matrix for the k-th case. k The inverse matrix; φ k β represents the regression vector of the k-th case; k v represents the adaptive gain step size factor for the k-th case; k This indicates the direction of the update of the influence factor for the k-th case.

[0095] Furthermore, the initial influence factor θ0 and the positive definite adaptive gain matrix P0 can be arbitrarily selected.

[0096] Furthermore, the update direction v of the influence factor for the k-th case. k The expression is:

[0097]

[0098] in, This represents the saturation function for the type of punishment classification of the k-th case.

[0099] Furthermore, the adaptive gain matrix P of the k-th case k The expression is:

[0100]

[0101] Among them, P k-1 Let a represent the adaptive gain matrix for the (k-1)th case; k β represents the adaptive gain for the k-th case; k This represents the adaptive gain step size factor for the k-th case;

[0102] Furthermore, the adaptive gain a for the k-th case k The expression is:

[0103]

[0104] Where, μ k This represents the step size parameter for the k-th case.

[0105] Furthermore, the adaptive gain step size factor β for the k-th case k The expression is:

[0106]

[0107] Among them, I [·] It is an indicator function; Δ k express c (i-1)k This represents the threshold for classifying the (i-1)th type of crime in the k-th case.

[0108] Furthermore, the fixed-term imprisonment prediction module is a saturated sentence prediction model based on sentencing logic. It utilizes the judgment features related to sentencing and case type output by the legal element identification unit enhanced by the large model, combined with a comprehensive assessment of the defendant's degree of social harm based on the substantive requirements of probation, to obtain a predicted value for the actual sentence. The fixed-term imprisonment prediction module of this invention sequentially confirms the role of criminal facts and sentencing characteristics in sentencing according to judicial practice, particularly considering the priority application of sentencing features related to conviction stipulated by law.

[0109] The expression for the fixed-term imprisonment sentence prediction module is:

[0110]

[0111] in, This represents the length of the sentence for the k-th case; This represents the saturation function of the fixed-term imprisonment prediction module for the k-th case; This serves as the starting point for sentencing in the k-th case; b represents the nth sentencing factor affecting the base sentence in the kth case; n m1 represents the unknown weight parameter corresponding to the nth sentencing factor; m2 represents the total number of sentencing factors affecting the base sentence; m3 represents the number of prior sentencing features stipulated in the law; and m4 represents the number of other unconsidered features. p represents the j-th prior sentencing characteristic stipulated in the sentencing guidelines in the k-th case. j For the unknown weight parameters corresponding to the j-th prior applicable sentencing feature; Let q represent the t-th unconsidered feature in the k-th case. tThe unknown weight parameters are for the t-th unconsidered feature; The bias term for the k-th case indicates that it is not included. and The combined influence of other possible quantitative factors.

[0112] Furthermore, the saturation function of the fixed-term imprisonment prediction module for the k-th case. The upper and lower limits are constituted by the upper and lower limits of the sentence for the corresponding crime as stipulated by law (such as the Criminal Law of the People's Republic of China), and are adjusted according to the different levels and severity of the statutory sentencing circumstances to reflect the differences in punishment at each stage of the sentencing ladder.

[0113] Furthermore, this invention employs the Adam optimization algorithm to optimize parameter b. n ,p j and q t Update as follows: Learning rate η is set to 0.001, momentum decay rate β1 = 0.9, second moment decay rate β2 = 0.999, batch size = 128, and smoothing term ∈ = 10. -8 The Adam algorithm updates each parameter θ∈{b} in the following way: n ,p j ,q t}:

[0114] This invention uses adjustment factor b n p j and q t This reflects the positive and negative impacts of different circumstances on the sentence. The positive and negative values ​​of these parameters correspond to the "addition in the same direction" and "subtraction in the opposite direction" sentencing strategies in the "Guiding Opinions on Sentencing for Common Crimes (Trial Implementation)." That is, when a circumstance helps to mitigate punishment, the parameter takes a positive value; and when a circumstance aggravates punishment, the parameter takes a negative value. This design aligns with judicial sentencing logic, making the model's parameters highly interpretable and enabling it to accurately simulate and predict the actual impact of various circumstantial factors on the sentence during the sentencing process.

[0115] Furthermore, the probation term prediction module is constructed based on the preconditions for probation (such as the offender must be sentenced to detention or imprisonment of no more than 3 years to be eligible for probation), combined with the substantive conditions for probation, namely the severity of the crime, remorse, risk of recidivism, and impact on the community. The probation term prediction module includes a probation judgment unit and a probation period unit.

[0116] Specifically, the expression for the applicable probation assessment unit is:

[0117]

[0118] Among them, h k∈{0,1} indicates whether probation is applicable to the k-th case; I(·) represents the indicator function; θ represents the regression vector composed of the judgment features of the k-th case; * The factor representing the degree of influence of each judgment circumstance feature on the applicable probation judgment unit is a vector of unknown parameters to be estimated for the applicable probation judgment unit; ∈ k This represents the impact of unmodeled dynamic or random noise in the k-th case.

[0119] The probation assessment unit constructed in this invention combines a two-step adaptive identification algorithm (a method used in the same type of sentence prediction model) to assess the unknown parameter vector θ. * An estimate is made, and a prediction is further given as to whether the defendant is eligible for probation.

[0120] Specifically, based on the substantive requirements of probation and the requirements of sentence duration, a saturated nonlinear probation period unit with a variable lower bound is constructed, expressed as follows:

[0121]

[0122] in, The probation period for the kth case (in months) This is the saturation function of the probation prediction module for the k-th case. Its lower saturation bound is the predicted fixed-term imprisonment value plus 1 year, and its upper saturation bound is 5 years. The input feature vector includes the predicted fixed-term imprisonment value obtained from the previous module. and the substantive requirements of probation ε is the influence factor of the probation period unit. k The k-th case is unmodeled dynamic or random noise.

[0123] Based on the constructed probation term prediction model, this invention uses the SGD optimization algorithm for parameter estimation and updating, where the learning rate is set to 0.001, the momentum is 0.9, and the batch size is 128.

[0124] Furthermore, this invention will also utilize the substantive elements of probation. t and probation period A predictive model for assessing the social danger level of the defendant is constructed, with the following expression:

[0125]

[0126] Where, γ k This serves as an evaluation index for the degree of social danger posed by the defendant in the k-th case. Let be the saturation function constructed based on the probation period requirement for the k-th case, with a lower saturation bound of 0.2 and an upper saturation bound of 1. ρ represents the substantive elements of probation, ρ is the influence factor of each judgment circumstance on the predictive model for assessing the defendant's social dangerousness, and ω is the unknown parameter vector to be estimated in the predictive model for assessing the defendant's social dangerousness; k The k-th case is unmodeled dynamic or random noise.

[0127] Based on the constructed model, this invention uses the NAG optimization algorithm for parameter estimation and updating, with the learning rate set to 0.001, momentum to 0.9, and batch size to 128.

[0128] This invention addresses the newly given substantive requirements for probation. Based on the estimated values ​​of the unknown parameter vectors obtained from the first k-1 cases, the model can output a predicted value of the defendant's social danger level. This predicted value can be used as an input feature for a fixed-term imprisonment prediction module to improve the accuracy of fixed-term imprisonment prediction. k The prediction accuracy.

[0129] Specifically, the regional difference handling module adjusts the unknown parameter vectors in three models—crime type classification, fixed-term imprisonment sentence prediction, and probation sentence prediction—based on the differences in legal practices across different provinces, ensuring that the model can adapt to the legal environment of different regions. The method is as follows: First, the model is trained on the training set of all provinces nationwide, and the parameters with the highest accuracy on the test set are selected as the initial parameters of the national model. Then, based on the national model parameters, fine-tuning is performed on the training sets of different provinces, and the model is tested on the test set of each province. The parameters with the highest accuracy are saved as the final model parameters for that province. In this way, the model maintains universality nationwide while being optimized for the characteristics of legal practices in each province. In the judicial sentencing process, different regions may have different legal practices and judgment standards. These differences in the application of law between regions often stem from local regulations, judicial interpretations, court judgment practices, and local social culture. Therefore, a single nationally unified model may not fully consider the actual situation of each province, leading to deviations in sentence prediction in some regions. To effectively address this challenge, this system is designed with a regional difference processing module, which aims to dynamically adjust model parameters based on the differences in legal practices in different provinces, ensuring that the model can accurately adapt to the legal environment and judicial practices in various regions.

[0130] The judgment output display module is used to output the type of sentence, sentence duration, and probation period.

[0131] Ultimately, the judgment output display module generates the type of sentence, probation, and specific duration of imprisonment based on the prediction results, and clearly displays the prediction results to the user. This module can present the sentencing results through an appropriate interface, ensuring that judicial personnel can clearly understand and apply the prediction results to make reasonable judicial decisions. Through the implementation of this system, accurate sentencing predictions can be provided to judicial personnel, ensuring the fairness and transparency of the judiciary and supporting the adaptation of sentencing standards across different regions. The system can also be flexibly adjusted according to the characteristics of different cases, demonstrating the innovation and practical application value of this invention.

[0132] To illustrate the effectiveness of the method proposed in this invention, let's take the crime of causing a traffic accident as an example. Based on the characteristics selected according to the "Criminal Law of the People's Republic of China" and the "Guiding Opinions on Sentencing for Common Crimes," the characteristics for the crime of causing a traffic accident are categorized into six aspects: elements of the crime, mitigating circumstances, aggravating circumstances, prior sentencing circumstances, statutory mitigating circumstances, and other characteristics. Elements of the crime include the number of deaths, the number of serious injuries, the escape, and the division of responsibility for the accident; mitigating circumstances include force majeure and first-time offenders; aggravating circumstances include repeat offenders and drunk driving; prior sentencing circumstances involve different age groups and mentally ill individuals; statutory mitigating circumstances include surrender and meritorious service; other characteristics include case number and province. Ultimately, these characteristics are linked to the type of punishment, the term of imprisonment, and the probation period in the judgment.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A data-driven sentence prediction system, characterized in that, It includes an input display module, a legal judgment element extraction module, a fixed-term imprisonment prediction module, a probation term prediction module, and a judgment output display module; The input display module is used to input and display judgment documents and case location information; The legal judgment element extraction module includes a document structure splitting unit and a large-model-enhanced legal element identification unit. The document structure splitting unit divides the judgment document into the facts ascertained during the trial and the court's opinion. The large-model-enhanced legal element identification unit performs legal provision matching, keyword matching, and large language model extraction based on the key case factors and sentencing factors in the facts ascertained during the trial and the court's opinion to obtain judgment details related to sentencing and case type. Among them, the legal provision matching method determines whether the defendant has the circumstances stipulated in the legal provision based on the legal provisions cited in the judgment document; Keyword matching uses positive-side matching, while regular expression matching is used to extract structured and meaningful plot features. Large language models are used to extract and determine plot features; The fixed-term imprisonment prediction module is based on the judgment characteristics related to sentencing and case type output by the legal element identification unit enhanced by the large model, combined with the comprehensive assessment of the defendant's degree of social harm given based on the substantive requirements of probation, to obtain the predicted sentence value. The probation term prediction module is constructed based on the preconditions for the application of probation and the substantive conditions of probation, and is used to determine the predicted probation value. The judgment output display module generates the duration of probation and fixed-term imprisonment based on the prediction results and displays the prediction results; the prediction results include the predicted sentence and the predicted probation.

2. The data-driven sentence prediction system according to claim 1, characterized in that, The characteristics of the circumstances for determining the circumstances include the elements constituting the crime, the circumstances applicable to sentencing prior to the crime, the statutory circumstances for mitigating the punishment, the circumstances for lenient punishment, and the circumstances for aggravating the punishment.

3. The data-driven sentence prediction system according to claim 1, characterized in that, The substantive conditions for probation include the severity of the crime, remorse, risk of recidivism, and impact on the community.

4. The data-driven sentence prediction system according to claim 1, characterized in that, The probation term prediction module includes a probation assessment unit and a probation period unit.

5. The data-driven sentence prediction system according to claim 4, characterized in that, It also includes a regional difference processing module, which adjusts the parameter vectors in the fixed-term imprisonment term prediction module and the probation term prediction module according to the differences in legal practices in different provinces, so that the model can adapt to the legal environment of different regions.

6. The data-driven sentence prediction system according to claim 1, characterized in that, The expression for the fixed-term imprisonment sentence prediction module is: in, This represents the length of the sentence for the k-th case; This represents the saturation function of the fixed-term imprisonment prediction module for the k-th case; This serves as the starting point for sentencing in the k-th case; b represents the nth sentencing factor affecting the base sentence in the kth case, where n = 1, 2, ..., m1; n m1 represents the weighting parameter corresponding to the nth sentencing factor; m2 represents the total number of sentencing factors affecting the base sentence; m3 represents the number of prior sentencing features stipulated in the law; and m4 represents the number of other unconsidered features. p represents the j-th prior sentencing feature specified in the sentencing guidelines for the k-th case; j For the unknown weight parameters corresponding to the j-th prior applicable sentencing feature; q represents the t-th unconsidered feature in the k-th case; t Let be the weight parameter for the t-th unconsidered feature; This is the bias term for the k-th case; Represents a real number.

7. The data-driven sentence prediction system according to claim 1, characterized in that, It also includes a crime type determination unit, which is used to determine the type of crime that the defendant should be punished based on the obtained crime characteristics, thereby obtaining a crime type classification.

Images