CN116823541A - Criminal investigation calculation method and system based on nonlinear model - Google Patents

Abstract

The invention relates to the technical field of legal text processing, in particular to a criminal calculation method and system based on a nonlinear model, comprising the following steps: text data in case description is obtained, and the text data are preprocessed to obtain sentencing factor characteristics; determining a criminal starting point based on the obtained criminal factor characteristics, and determining a range of a criminal range of the criminal rule adjustment reference according to the criminal factor characteristics to obtain the weight of the criminal rule; estimating noise of the nonlinear model, determining noise distribution, generating a plurality of samples obeying the noise distribution based on the sentencing factor characteristics, and determining error limits of parameter estimation values of the nonlinear model through a plurality of simulations; and in the error limit, according to the corresponding relation between the weight of the sentencing factors and the parameter estimation value, determining the estimation value and the confidence limit of the unknown parameter vector corresponding to the sentencing characteristic factors in the nonlinear model, and obtaining the prediction criminal period output by the model as a reference standard for judgment.

Description

A sentencing calculation method and system based on nonlinear models

Technical field

The invention relates to the technical field of legal text processing, specifically a sentencing calculation method and system based on a nonlinear model.

Background technique

The statements in this section merely provide background technical information related to the present invention and do not necessarily constitute prior art.

The sentencing system is a text data processing system that outputs predictions of sentencing results to judicial workers based on text information in case files. This type of system usually relies on features extracted from a large amount of text data in case files, and based on the impact of sentencing factors on sentencing outcomes in different cases , and the results are output after certain processing and calculations. Currently, this type of system is divided into two categories. One is a system built based on traditional linear regression statistical models or empirical methods. The other is a system based on machine learning and natural language processing technology to mine key information in the case file text and the decisions contained in the text. logic.

Among them, the linear regression statistical model uses statistical methods to process the original text data in the case file according to the results of the law of large numbers and the central limit theorem to obtain the sentencing calculation results. However, the model itself does not limit the sentence range and cannot It is fully adapted to the actual sentencing scenario (the actual sentencing scenario has non-linear saturation characteristics), so the sentencing mechanism contained in the text data cannot be accurately analyzed and controlled, making the sentencing results obtained more different from the sentencing scenario described in the actual case. It is difficult to help judicial staff improve their work efficiency. In addition, due to the limited number of currently disclosed case files, statistical methods need to assume a priori that the data satisfies good statistical properties (such as independent and identical distribution, etc.), making it difficult to determine the values and accuracy of the estimated parameters in the model, making the model impossible to master. The adjustment ratio of the sentencing circumstances to the base sentence in different cases.

For related technologies and methods of machine learning and natural language processing, a large amount of legal text information is usually required to train a model with strong generalization capabilities, and the number of currently public case files (judgment text) cannot support the model to obtain reliable output. ability, making it difficult to determine the accuracy of estimated parameters in the model through training.

Contents of the invention

In order to solve the technical problems existing in the above background technology, the present invention provides a sentencing calculation method and system based on a nonlinear model. Based on the effective information extracted from the case file text, the nonlinear saturation model is applied, and Bayesian embedding and random simulation are used. With the multi-stage calculation method, it solves the problem of determining the accuracy of parameter estimation in the process of obtaining sentencing results under limited data samples. It overcomes the limitation of existing methods that require sufficient big data sample size, and also shows the impact of various sentencing characteristics. Trends over time.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A first aspect of the present invention provides a sentencing calculation method based on a nonlinear model, which includes the following steps:

Obtain the text data in the case description and pre-process it to obtain the sentencing factor characteristics;

Determine the starting point of sentencing based on the characteristics of the sentencing factors obtained, determine the range of the sentencing circumstances to adjust the range of the benchmark penalty based on the characteristics of the sentencing factors, and obtain the weight of the sentencing factors;

Estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limits of the nonlinear model parameter estimates through several simulations;

Within the error bound, based on the correspondence between the weights of sentencing factors and parameter estimates, determine the estimated values and confidence limits of the unknown parameter vector corresponding to the sentencing characteristic factors in the nonlinear model, and obtain the predicted sentence output of the model as Reference base for adjudication.

Preprocessing includes: extracting sentencing-related feature fields from the text data and merging them to obtain structured case text data.

The nonlinear model is a saturated nonlinear regression model, and the floating upper and lower bounds of the saturated nonlinear regression model are determined according to the case type.

Estimating the noise of the nonlinear model specifically includes: obtaining the estimated noise of the nonlinear model based on the least squares method, using the estimated noise to obtain the empirical distribution curve, and determining the normal density function and variance of the noise.

Determine the starting point of sentencing, specifically: Divide the sentencing interval into several equal parts, calculate the accuracy and estimated value of the bias term for each decimal respectively, and on the premise of ensuring that the calculation accuracy meets the set value, minimize the bias term and determine the starting point of sentencing. position within the sentencing range.

Obtain the weight of the sentencing factors, specifically: determine the weight of the sentencing factors based on a multi-stage stochastic quasi-Newton adaptive learning algorithm.

After several simulations, the error bounds of the estimated values of the nonlinear model parameters are determined, specifically:

Obtain several preprocessed text data as samples, and obtain a multidimensional output observation set based on the nonlinear model;

The multi-dimensional output observation set is based on the multi-stage stochastic quasi-Newton adaptive learning algorithm to obtain parameter estimates for the corresponding number of simulations;

According to the parameter estimation error of a certain dimensional component and the empirical distribution function, it is determined that the corresponding parameter estimation error belongs to the interval where the error limit is at least with a certain probability.

A second aspect of the present invention provides a system required to implement the above method, including:

The text preprocessing module is configured to: obtain the text data in the case description, and preprocess to obtain the sentencing factor characteristics;

The first parameter estimation module is configured to: determine the sentencing starting point based on the obtained characteristics of the sentencing factors, determine the range of the sentencing circumstances to adjust the range of the benchmark punishment based on the characteristics of the sentencing factors, and obtain the weight of the sentencing factors;

The second parameter estimation module is configured to: estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limit of the nonlinear model parameter estimate through several simulations;

The result output module is configured to: within the error limit, determine the estimated value and confidence limit of the unknown parameter vector corresponding to the sentencing characteristic factor in the nonlinear model based on the correspondence between the weight of the sentencing factor and the parameter estimate. , the predicted sentence output from the model is obtained as a reference benchmark for judgment.

A third aspect of the invention provides a computer-readable storage medium.

A computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the sentencing calculation method based on a nonlinear model are implemented as described above.

A fourth aspect of the invention provides a computer device.

A computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a sentencing calculation based on a nonlinear model as described above. steps in the method.

Compared with the existing technology, one or more of the above technical solutions have the following beneficial effects:

1. Determine the starting point of sentencing and the weight of sentencing factors based on the effective information extracted from the text of the real case file, and then determine the error limits of the model parameter estimates through multiple simulations based on the generated sample data. Through the size and parameters of the weight of sentencing factors The correspondence between the estimated values can obtain the value of the unknown parameter vector corresponding to the sentencing characteristic factors in the nonlinear model, so that when calculating the sentence, the model can better grasp the relationship between sentencing circumstances in different cases under limited data samples. Adjust the proportion of the base sentence to better adapt to actual sentencing scenarios.

2. The sentencing scenarios in each case description are independent. The adjustment ratio of the sentencing circumstances to the base penalty in different cases is different. It is also difficult to find multiple identical sentencing scenarios to train the model. However, based on real case descriptions The "weight estimate" given by the text data in the text data is then used to give the "weight estimate" given by the randomly generated data, and the correspondence between the two parts of data is used to obtain the estimated high probability confidence bound, so that the model can calculate the sentence period. With limited data samples, we can better understand the adjustment ratio of sentencing circumstances to the base sentence in different cases.

3. The nonlinear saturation model can be well adapted to sentencing scenarios and can limit the announced sentence within the required statutory sentencing range for cases exceeding or falling below the corresponding statutory sentencing range to make up for the applicability limitations of the traditional linear model. , and can adapt to the analysis needs of small data samples.

4. It can reliably estimate sentencing characteristics and present the changing trend of the influence of each sentencing characteristic over time, which is helpful to help judicial staff analyze and determine the sentencing characteristic factors corresponding to the case, thereby ensuring that different types of cases can output prediction results time reliability.

Description of the drawings

The description and drawings that constitute a part of the present invention are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

Figure 1 is a schematic diagram of the overall architecture of the sentencing calculation method provided by one or more embodiments of the present invention;

Figure 2 is a schematic diagram comparing the sentencing calculation accuracy of serious injury cases between the S-model and the L-model provided by one or more embodiments of the present invention;

Figure 3 is a schematic diagram of the change trend of the S-model part of the variables of interest provided by one or more embodiments of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and examples.

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

The basic steps of sentencing include determining the starting point of the sentencing period, determining the base sentence and determining the declaratory sentence. The starting point of the sentencing period is the sentencing period determined based on the basic facts in the description of the case, while the base sentence needs to be based on every fact that occurred during the description of the case as the actual situation. Adjustments are made so that the benchmark penalty is added to the starting point of the sentence to form a declaratory sentence, and the parameters in the description of the case that can affect the determination of the benchmark sentence are sentencing factors.

As described in the background art, when using the sentencing system to process case file text information and output sentencing results for reference, there are the following problems:

First, in terms of the applicability of the model, the traditional linear model lacks the restriction of the sentence interval and has limited applicability, while the deep learning model has a large demand for data volume.

Second, in terms of calculation methods, the theory of existing calculation methods requires that the data have strong statistical assumptions, and there is a lack of guarantee for the reliability of calculations with limited data samples.

Therefore, the following embodiments provide a sentencing calculation method and system based on a nonlinear model. Based on the effective information extracted from the case file text, the nonlinear saturation model is applied, and Bayesian embedding, stochastic simulation and multi-stage calculation methods are used to solve the problem. Under limited data samples, in the process of obtaining sentencing results, the problem of determining the accuracy of parameter estimation overcomes the limitation of simple statistical methods that require sufficient big data sample size. It also shows the changing trend of the impact of various sentencing characteristics over time.

Example 1:

As shown in Figures 1-3, a sentencing calculation method based on a nonlinear model includes the following steps:

Obtain the text data in the case description and pre-process it to obtain the sentencing factor characteristics;

Determine the starting point of sentencing based on the characteristics of the sentencing factors obtained, determine the range of the sentencing circumstances to adjust the range of the benchmark penalty based on the characteristics of the sentencing factors, and obtain the weight of the sentencing factors;

Estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limits of the nonlinear model parameter estimates through several simulations;

Within the error bound, based on the correspondence between the weights of sentencing factors and parameter estimates, determine the estimated values and confidence limits of the unknown parameter vector corresponding to the sentencing characteristic factors in the nonlinear model, and obtain the predicted sentence output of the model as Reference base for adjudication.

specific:

1 Text preprocessing

Step 1-1: For judicial judgment document information and case description text, natural language processing technology is used to extract sentencing-related feature fields based on reasonable segmentation and manual annotation. The obtained natural features are used as factors affecting sentencing.

Step 1-2: On this basis, further feature selection is performed according to different calculation purposes. The specific process is as follows:

Step 1-2-1: Classify according to the legal attributes of these fields (for example, merging "previous record" with "received criminal punishment for rape"), and several "natural characteristics" such as obtaining a suspended sentence and filing ancillary civil lawsuits are used for the sentence. calculate.

Step 1-2-2: Merge natural features based on relevance; some natural features are legally relevant or similar, and such features can be merged. For example, according to Article 26 of the Criminal Law, "principal offenders" include "chief criminals", "general principal offenders", and "hiring others". Therefore, the three natural characteristics of the above types can be combined into "principal offenders". For example, according to Article 67 of the Criminal Law, "surrender" includes "quasi-surrender", "active surrender", "persuasion to surrender" and other types. Therefore, the three natural characteristics of the above types can be combined into "surrender".

This embodiment combines the natural features extracted from the text, and deletes some features that are too sparse (for example, the frequency is less than 200, which is not enough to have an impact after adding them) according to the distribution of specific features. In order to further eliminate the influence of these characteristics, cases containing these characteristics were also deleted accordingly. Finally, the main characteristic factors are determined in the modeling.

2 Apply saturated nonlinear regression model

Step 2-1: Combine the common characteristics of sentencing scenarios to determine the floating upper and lower bounds of the saturated nonlinear regression model:

;

Among them, the specific definition of the saturation function is as follows:

;

In the model, y _t is the fixed-term imprisonment term of the t- th case (unit: month); the cases are sorted according to the time of judgment. Cases occurring on the same day can be sorted randomly, which has no substantial impact on the calculation results.

Indicates the starting point of sentencing (unit: month);/> In the crime of intentional injury, they respectively represent the number of minor injuries, the number of serious injuries, and the number of deaths that determine the amount of punishment in the t-th case (in the crime of counterfeiting registered trademarks, they respectively represent the amount of illegal business operations, the amount of illegal income, and the amount of illegal income involved in the t-th case. The number of trademark types, in the case of fund-raising fraud, respectively represents the amount of fraud in the t- th case, etc.);

In the crime of intentional injury , b, c, and d respectively represent the increased sentence for each additional minor injury, serious injury, or death (in the crime of counterfeiting registered trademarks and the crime of fund-raising fraud, respectively, they represent the degree of harm of the corresponding factors);

Represents the regression vector composed of selected sentencing characteristic factors;/> Represents the unknown parameter vector corresponding to the sentencing characteristic factors. Each component reflects the percentage of the corresponding characteristic factor; e is the bias term of the modeling, which represents the comprehensive impact of other sentencing characteristic factors that may not be considered;/> is the possible random noise; U _t is the upper limit of the corresponding statutory sentencing interval, and L _t is the lower limit of the corresponding statutory sentencing interval. They vary depending on the nature of the case and fluctuate according to the nature of the statutory sentencing circumstances.

3 Calculation and analysis methods

Step 3-1: Introduce new high-dimensional regression vectors as follows:

;

Accordingly, define the unknown parameter vector:

;

Correspondingly, the nonlinear stochastic model in step 2-1 can be transformed into the following saturated model:

;

in, is the known input vector, y _t is the output, and θ is the unknown parameter to be estimated. Usually according to the actual case scenario, the prior compact convex set D where the parameters are located can be given.

The actual meaning of the parameter θ is "the adjustment ratio of the sentencing circumstances to the base sentence", which is the weight of the sentencing factors, that is, the true value of the parameter θ . The projection interval is the range stated a priori of the parameter θ ; and the prior set D The scope corresponds to the specific provisions in the law regarding the adjustment of the base penalty range based on sentencing circumstances.

Step 3-2: In order to solve the problem of the distribution function of model noise required in specific sentencing calculations, first use the least squares method to estimate the noise of the specific sentencing model based on judicial judgment data.

Specifically, only samples in the non-saturated interval are selected. When the t-th sample Upon arrival, the iterative least squares algorithm is used to update the estimated value of parameter θ /> , and then use the estimated values of the parameters/> Give sample/> Noise estimation in/> , the specific calculation formula is as follows:

;

Estimation using noise The empirical distribution curve of noise is given, the normal density function and variance of noise are determined, and applied in the specific sentencing calculation process.

Step 3-3: In order to overcome the ill-posed difficulty in solving mathematical equations in the sentencing model, the uniform segmentation method and the prediction accuracy comparison method are used to specifically determine the starting point of sentencing. Divide the sentencing interval into six equal parts, take the starting point of sentencing as 1/6, 1/3, 1/2, 2/3, and 5/6 of the sentencing interval and the lowest and highest points of the interval for calculation, and obtain 7 different groups. The calculation accuracy and bias term estimate.

The principle of selecting the sentencing starting point is to reduce the offset term as much as possible while ensuring high calculation accuracy. In this embodiment, it is finally determined through experiments that the sentencing starting point is selected at 1/3 of the corresponding sentencing interval.

Step 3-4: Use the multi-stage stochastic quasi-Newton adaptive learning method (MSQN) to calculate the weight of the sentencing factors.

A priori compact convex set D for parameters, and a positive definite matrix , introduce the projection operator projection operator/> ,defined as:

;

Among them, norm Defined as/> Use the above projection operator.

The actual meaning of the parameter θ is "the adjustment ratio of the sentencing circumstances to the base sentence", which is the weight of the sentencing factors, that is, the true value of the parameter θ . The projection interval is the range stated a priori of the parameter θ ; and the prior set D The scope corresponds to the specific provisions in the law regarding the adjustment of the base penalty range based on sentencing circumstances.

For each time t , the algorithm inputs the regression vector ;Model output y _t , regularization factor μ _j,t (1≤j≤K), algorithm stage number K, parameter projection prior set D, noise distribution/> , algorithm initial value P _j,0 (1≤j≤K),/> (1≤j≤K).

The iterative estimation formula at time t is obtained based on the MSQN adaptive learning method. The MSQN method (multi-stage stochastic quasi-Newton adaptive learning method) itself is an existing technology and will not be described in detail in this embodiment.

Specifically in the sentencing data analysis of this embodiment, K=3 is taken; through the estimation of data noise in step 3-1, the noise distribution is taken as a normal distribution with a mean value of 0 and a standard deviation of 5;

In addition, the function G _t (·) can be specifically expressed as:

G _t (x)=U _t +(L _t -x)F(L _t -x)-(U _t -x)F(U _t -x)+25[f(L _t -x)-f(U _t -x)];

The expression of the derivative G´ _t (·) is:

G´ _t (x)=F(U _t -x)-(L _t -x);

Among them, F(·) and f(·) are the distribution function and probability density function of the normal distribution N(0, 25) respectively;

The projection interval of the parameter corresponding to the characteristic factor "number of serious injuries" is [0,40], the projection interval of the parameter corresponding to the characteristic factor "number of minor injuries" is [0,10], and the projection interval of the offset term is [-1,1]. Among the remaining features, the projection interval of the corresponding parameters for the penalty-increasing characteristic factors (such as holding weapons) is [-0.1,1], and the projection interval of the corresponding parameters for the penalty-reducing characteristic factors (such as surrender) is [-1,0.1]; μ _{j, t} =25, 1≤j≤3; for 1≤j≤3 Initial value P _j,0 =I,/> , n is the sample size.

The multi-stage method (MSQN) in this embodiment is different from the traditional linear least squares (RLS) method and is designed for sentencing systems with saturated properties. This algorithm does not require the regression vector composed of data to satisfy traditionally difficult statistical assumptions such as independent and identical distribution. It is more suitable for the characteristics of complex text information such as judicial decisions, and can assist judicial staff in judging the actual impact of various sentencing circumstances on sentencing decisions. Provide calculation support.

Step 3-5: Comprehensive use of Bayesian embedding, stochastic simulation and multi-stage calculation methods to provide parameter estimation accuracy (high probability confidence bound) under limited sentencing data samples, overcoming the problem that pure statistical methods require sufficient large data sample size limitations. Includes the following steps:

Step 3-5-1: Extract N samples from independent and identical distributions , obey the distribution/> , where U is the uniform distribution on the parameter prior set D, and F is the noise obeying the normal distribution (using the noise estimation distribution obtained in step 3-2, the sample here is generated using the noise distribution F, not extracted from the original data of).

Step 3-5-2: Using N samples {X ₁ , X ₂ ,...X _N }, features And the saturated nonlinear regression model form generates an n-dimensional output observation set {Y ₁ , Y ₂ ,...Y _N }.

Step 3-5-3: Use the MSQN method to give parameter estimates for N simulations, and calculate the parameter estimation error of the j-th dimension component of the parameter (j=1,...m) , and then use the following formula to calculate/> The empirical distribution function of:

;

Step 3-5-4: For any and satisfy/> The j-th dimension parameter estimation error calculated using MSQN/> At least with probability/> Belongs to the following range:

;

in, is the empirical distribution/> of/> Quantile.

Step 3-6: Use step 3-1 with/> The corresponding relationship is given in the original model/> The estimation and estimation accuracy (high probability confidence bound).

Steps 3-4 are "weight estimates" given based on text data in real case descriptions. Here are "weight estimates" given using randomly generated data, which are used to obtain estimated high-probability confidence bounds so that the model can calculate When it comes to sentencing, it is possible to better understand the adjustment ratio of sentencing circumstances to the base sentence in different cases under limited data samples.

4 Verification

This example uses the crime of intentional injury as an example to conduct sentence prediction and factor analysis. The samples are taken from the first trial judgment data of the crime of intentional injury published between January 2011 and June 2021, totaling 199,590.

Step 4-1: Based on the nonlinear recursive identification algorithm in step 3-3, specific parameter estimates and their error bounds can be obtained based on the structured data extracted from the actual judgment document information, based on the S-model. The embodiment proposes a nonlinear saturation model, and the L-model is a traditional linear model.

Step 4-2: Use the nonlinear saturation model (MSQN algorithm, referred to as S-model) proposed in this embodiment and the parameter estimates calculated in step 4-1 to calculate the sentence period, and compare it with the traditional linear model (RLS algorithm , referred to as L-model), the specific results are shown in Table 1:

Table 1: Comparison of calculation accuracy between S-model and L-model

The sentencing calculation accuracy is defined as the average of the relative calculation errors (Prediction accuracy). It can be seen that the S-model has improved accuracy compared with the L-model in calculating the sentence for intentional injury causing serious injury.

Figure 2 shows the changing trend in the accuracy of sentencing calculations for serious injury cases from 2011 to 2021. It can be seen that the calculation accuracy of the S-model is always much higher than that of the L-model.

Step 4-3: In addition, the proposed MSQN algorithm can also be used to present the changing trend of the influence of sentencing characteristic elements over time, and then discover the changing rules of justice and the trajectory of changes in the rule of law behind it. Through analysis, it was found that most of the sentencing characteristic factors in the modeling showed stable changes over time, but there were also individual features that changed significantly. For example, the two features of bias items and confession and punishment showed significant changes over time. The specific changes are as shown in the figure 3 shown.

This embodiment enables the sentencing system to assist judicial decisions based on sentencing patterns extracted from legal texts and reliable document information, and output sentencing results to judicial staff for reference.

In terms of model applicability, a nonlinear saturated model with accuracy and interpretability is applied. This model can adapt well to sentencing scenarios, ensuring that the sentences for cases exceeding or falling below the corresponding legal sentencing range are limited to the statutory sentencing range to make up for the applicability limitations of the traditional linear model, and can be adapted to the analysis of small data samples need.

In terms of calculation methods, considering that complex social data such as judicial judgment texts are far from satisfying traditional data assumptions such as independent and identical distribution, adaptive methods are proposed to establish theoretical guarantees under weaker data conditions.

In terms of accuracy guarantee, the reliability guarantee of parameter estimation in limited data sample situations is theoretically provided to accurately define the actual effect of sentencing circumstances.

In terms of computational effects. Compared with the traditional linear regression model, the established nonlinear model and the corresponding new calculation method can provide interpretable sentencing results based on the given case description, and the accuracy of the sentencing calculation has been improved (with intentional injury) crime as an example), which can provide the judge with more valuable sentencing recommendations.

Example 2:

Systems that implement the above methods include:

The text preprocessing module is configured to: obtain the text data in the case description, and preprocess to obtain the sentencing factor characteristics;

The first parameter estimation module is configured to: determine the sentencing starting point based on the obtained characteristics of the sentencing factors, determine the range of the sentencing circumstances to adjust the range of the benchmark punishment based on the characteristics of the sentencing factors, and obtain the weight of the sentencing factors;

The second parameter estimation module is configured to: estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limit of the nonlinear model parameter estimate through several simulations;

The result output module is configured to: within the error limit, determine the estimated value and confidence limit of the unknown parameter vector corresponding to the sentencing characteristic factor in the nonlinear model based on the correspondence between the weight of the sentencing factor and the parameter estimate. , the predicted sentence output from the model is obtained as a reference benchmark for judgment.

Embodiment three:

This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of a sentencing calculation method based on a nonlinear model as described in the first embodiment are implemented.

Embodiment 4:

This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in the first embodiment above. The steps in a sentencing calculation method based on a nonlinear model.

Each step or network involved in the above Embodiments 2 to 4 corresponds to Embodiment 1. For specific implementation details, please refer to the relevant description of Embodiment 1. The term "computer-readable storage medium" shall be understood to include a single medium or multiple media that includes one or more sets of instructions; and shall also be understood to include any medium capable of storing, encoding, or carrying instructions for use by a processor. The executed instruction set causes the processor to perform any method in the present invention.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. A sentencing calculation method based on a nonlinear model, which is characterized by including the following steps: Obtain the text data in the case description and pre-process it to obtain the sentencing factor characteristics; Determine the starting point of sentencing based on the characteristics of the sentencing factors obtained, determine the range of the sentencing circumstances to adjust the range of the benchmark penalty based on the characteristics of the sentencing factors, and obtain the weight of the sentencing factors; Estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limits of the nonlinear model parameter estimates through several simulations; Within the error bound, based on the correspondence between the weights of sentencing factors and parameter estimates, determine the estimated values and confidence limits of the unknown parameter vector corresponding to the sentencing characteristic factors in the nonlinear model, and obtain the predicted sentence output of the model as Reference base for adjudication. 2. A sentencing calculation method based on a nonlinear model according to claim 1, characterized in that the preprocessing includes: extracting and merging characteristic fields related to sentencing in the text data to obtain structured case text data. . 3. A sentencing calculation method based on a nonlinear model as claimed in claim 1, characterized in that the nonlinear model is a saturated nonlinear regression model, and the floating upper and lower bounds of the saturated nonlinear regression model are determined according to the case type. 4. A sentencing calculation method based on a nonlinear model as claimed in claim 1, characterized in that the noise of the nonlinear model is estimated, specifically: obtaining the estimated noise of the nonlinear model based on the least squares method, and using the estimated noise to obtain The empirical distribution curve determines the normal density function and variance of the noise. 5. A sentencing calculation method based on a nonlinear model as claimed in claim 1, characterized in that determining the starting point of sentencing is specifically: dividing the sentencing interval into several equal parts, and calculating the accuracy and offset terms of each equal part respectively. The estimated value, on the premise of ensuring that the calculation accuracy meets the set value, minimizes the offset term and determines the position of the sentencing starting point in the sentencing interval. 6. A sentencing calculation method based on a nonlinear model as claimed in claim 1, characterized in that the weight of the sentencing factors is obtained, specifically: determining the weight of the sentencing factors based on a multi-stage random quasi-Newton adaptive learning algorithm. . 7. A sentencing calculation method based on a nonlinear model as claimed in claim 1, characterized in that the error limits of the nonlinear model parameter estimates are determined through several simulations, including: Several preprocessed text data are obtained as samples, and a multidimensional output observation set is obtained based on the nonlinear model. 8. A sentencing calculation method based on a nonlinear model as claimed in claim 7, characterized in that the error limit of the nonlinear model parameter estimate is determined through several simulations, and further includes: The multi-dimensional output observation set is based on a multi-stage stochastic quasi-Newton adaptive learning algorithm to obtain parameter estimates for the corresponding number of simulations. 9. A sentencing calculation method based on a nonlinear model as claimed in claim 8, characterized in that the error limit of the nonlinear model parameter estimate is determined through several simulations, and further includes: According to the parameter estimation error of a certain dimensional component and the empirical distribution function, it is determined that the corresponding parameter estimation error belongs to the interval where the error limit is at least with a certain probability. 10. A sentencing calculation system based on a nonlinear model, characterized by: The text preprocessing module is configured to: obtain the text data in the case description, and preprocess to obtain the sentencing factor characteristics; The first parameter estimation module is configured to: determine the starting point of sentencing based on the obtained characteristics of sentencing factors, determine the range of the sentencing circumstances to adjust the range of the benchmark punishment based on the characteristics of sentencing factors, and obtain the weight of sentencing factors; The second parameter estimation module is configured to: estimate the noise of the nonlinear model and determine the noise distribution, generate a number of samples that obey the noise distribution based on the characteristics of the sentencing factors, and determine the error limit of the nonlinear model parameter estimate through several simulations; The result output module is configured to: within the error limit, determine the estimated value and confidence limit of the unknown parameter vector corresponding to the sentencing characteristic factor in the nonlinear model based on the correspondence between the weight of the sentencing factor and the parameter estimate. , the predicted sentence output from the model is obtained as a reference benchmark for judgment.