Case auxiliary analysis method and system based on saturated neural network

Abstract

The present disclosure provides a case auxiliary analysis method and system based on a saturated neural network, relates to the technical field of case auxiliary analysis, obtains case description text and judgment document information, automatically extracts feature fields, performs special element smoothing processing after obtaining the feature fields, and constitutes structured nonlinear data; the structured nonlinear data is input into a saturated neural network model, a sentencing time estimation value of the case is calculated, specific crime sentencing interval common attribute characteristics are combined, a saturated restriction with a floating upper and lower limit is nested in the output layer of the saturated neural network model, and a back propagation algorithm based on the saturated restriction and an average relative error is used to take the relative error as an optimization target of the network, derive a step length update of each parameter of the network by using a gradient term, adaptively update a parameter estimation value, and output an auxiliary case analysis result until the output auxiliary case analysis result. The present disclosure makes up for the applicable limitations of traditional machine learning models and linear regression models lacking interpretability.

Description

Technical Field

[0001] This disclosure relates to the field of case analysis technology, specifically to a case analysis method and system based on saturated neural networks. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Currently, the key objectives of intelligent sentencing calculation systems based on judgment data are twofold: first, to efficiently extract effective case features from case description texts; and second, to accurately output relatively accurate sentence calculation results for different types of cases, rather than artificially prescribed sentence ranges or intervals.

[0004] For the former, accurately extracting the characteristics of the judgment is crucial yet challenging. Currently, the main technical approach involves using word vector models to train on a large amount of document data to uncover logical connections and implicit features within legal texts. However, in actual legal document data, the amount of case data varies significantly depending on the specific circumstances, and the main factors influencing sentencing are not uniform. Cases with fewer than 10,000 cases account for 83.6% of all cases. Given this small sample size, simply using word vector models is insufficient to accurately capture the key information affecting sentencing for different case scenarios.

[0005] For the latter, there are currently two main approaches: First, there are traditional machine learning techniques and methods. While these have received widespread attention and application in assisting sentencing calculations and have shown significant advantages in sentencing accuracy, existing methods remain "black box models," failing to provide support for the interpretability and reliability of the results and algorithms. The interpretability of the underlying mathematical model is crucial. Second, there are classic linear regression models and calculation methods. These models offer some interpretability in their mechanisms, but they do not restrict the sentencing range and cannot fully adapt to actual judgment data with nonlinear saturation characteristics. Therefore, it is difficult to achieve in-depth analysis and accurate understanding of the inherent mechanisms of assisting sentencing calculations. Furthermore, in terms of assisting sentencing accuracy, the computational performance of linear regression models is still significantly inferior to that of machine learning algorithms. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes a case-assisted analysis method and system based on saturated neural networks. It utilizes saturated neural network methods for auxiliary calculations in case analysis and proposes an element extraction method combining large language models and regular expressions for text preprocessing. This enables accurate analysis and rapid calculation of various types of cases, providing legal professionals with more valuable auxiliary analysis suggestions.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] Case-assisted analysis methods based on saturated neural networks include:

[0009] The system acquires case description text and judgment documents, performs text preprocessing, automatically extracts feature fields, and performs special element smoothing processing after acquiring feature fields to form structured nonlinear data.

[0010] Structured nonlinear data is input into a saturated neural network model to assist in calculating the estimated sentencing time for a case. Combining the common attribute characteristics of the sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain the auxiliary case analysis results.

[0011] According to some embodiments, the present disclosure adopts the following technical solutions:

[0012] A case-aided analysis system based on saturated neural networks includes:

[0013] The data acquisition module is used to acquire case description text and judgment document information, and to perform text preprocessing, automatically extract feature fields, and perform special element smoothing after acquiring feature fields to form structured nonlinear data.

[0014] The case analysis quantification module is used to input structured nonlinear data into a saturated neural network model to assist in calculating the estimated sentencing time for a case. Combining the common attribute characteristics of the sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain auxiliary case analysis results.

[0015] According to some embodiments, the present disclosure adopts the following technical solutions:

[0016] A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned case-assisted analysis method based on a saturated neural network.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the case-assisted analysis method based on a saturated neural network.

[0019] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0020] This disclosure proposes a case-aided analysis method based on saturated neural networks. It presents a saturated neural network model that is both accurate and interpretable. This model ensures that the sentencing information for cases exceeding or falling below the corresponding statutory sentencing range is limited to that range. The output layer of the saturated neural network model incorporates floating upper and lower bounds for saturation, overcoming the limitations of traditional machine learning and linear regression models in terms of interpretability, while retaining the computational advantages of machine learning methods in sentencing accuracy. Furthermore, it is worth noting that the saturated neural network model proposed in this disclosure is universally applicable to judicial sentencing calculation problems. Depending on the specific problem scenario, the feedforward neural network can be replaced with any more complex or practical network structure.

[0021] In terms of algorithm design, this disclosure takes into account the nonlinear characteristics of complex social data such as judicial judgment texts, and proposes a backpropagation algorithm based on saturation constraints and average relative error (SC-MRE-BP). This algorithm follows the saturation properties of the data itself, and unifies the L1 norm relative error formula to be optimized and the precision formula for measuring sentencing accuracy, so as to ensure that the estimated sentence can converge to the true sentence value more quickly.

[0022] This disclosure achieves fully automatic, efficient, and accurate text preprocessing by combining the ERNIE Bot 4.0 large language model with regular expressions. The average extraction accuracy can reach over 90%, overcoming the shortcomings of classic word vector models in accurately capturing the main feature factors affecting sentencing when the sample size is small.

[0023] In terms of accuracy, compared to traditional neural network models and linear regression models, the proposed saturated neural network model and the corresponding SC-MRE-BP algorithm can calculate interpretable sentencing results based on given document descriptions, with accuracy improvements of approximately 1.01% and 8.56%, respectively. Furthermore, this invention achieves sentencing accuracy of over 85% even with small sample sizes, demonstrating that, considering judicial discretion, this sentencing calculation method can provide valuable sentencing recommendations based on only a small amount of data to meet the needs of assisting judges in sentencing. Attached Figure Description

[0024] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0025] Figure 1 This is a diagram illustrating the overall architecture of the method according to an embodiment of this disclosure;

[0026] Figure 2 This is a schematic diagram of the structure of a saturated neural network model according to an embodiment of the present disclosure;

[0027] Figure 3 This is a flowchart of the SC-MRE-BP algorithm according to an embodiment of the present disclosure. Detailed Implementation

[0028] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, 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 disclosure pertains.

[0030] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] Example 1

[0032] One embodiment of this disclosure provides a case-assisted analysis method based on a saturated neural network, including:

[0033] Step 1: Obtain the case description text and judgment document information, and perform text preprocessing, automatically extract feature fields, and perform special element smoothing after obtaining feature fields to form structured nonlinear data;

[0034] Step 2: Input structured nonlinear data into a saturated neural network model to assist in calculating the estimated sentencing time for the case. Combining the common attribute characteristics of the specified sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain the auxiliary case analysis results.

[0035] The auxiliary case analysis results include sentencing estimates and sentencing estimates for similar historical cases.

[0036] As one example, the specific implementation process of the case-assisted analysis method based on saturated neural networks is as follows:

[0037] Step 1-1: For judicial judgment documents and case description texts, based on reasonable segmentation of different modules, a method combining large language models and regular expressions is used to efficiently extract sentencing-related feature fields. The specific technical process is as follows:

[0038] Step 1-1-1: For semantically complex elements that play a decisive role in case analysis (penalty sentencing) and for which accurate text recognition rules cannot be established, Baidu Wenxin Yiyan (ERNIE Bot 4.0) is used to achieve fully automatic and accurate extraction. For example, extracting elements such as the amount of money involved and the number of people involved from the text. This disclosure achieves fully automatic and accurate extraction based on Baidu Wenxin Yiyan (ERNIE Bot 4.0), with an extraction accuracy rate of over 90%.

[0039] Step 1-1-2: For elements in the case that are semantically simple and have clearly defined text recognition rules, regular expressions are used to achieve efficient extraction. For example, elements such as confession and surrender in the text can be efficiently extracted using regular expressions, with the accuracy rate of main feature extraction reaching over 90%.

[0040] Steps 1-2: Based on this, and according to different computational purposes and sample distributions, the decision features are further processed to obtain the final structured data. The specific technical process is as follows:

[0041] Step 1-2-1: Based on the distribution of specific features, some excessively sparse features (frequency less than 0.1%, insufficient to have an impact after addition) were removed. To further eliminate the influence of these features, cases containing these features were also removed accordingly, ultimately determining the main feature factors in the modeling.

[0042] Step 1-2-2: For judgment factors (such as the amount of fraud) that fluctuate significantly and vary greatly in different cases, use quantile grading (usually using 20%, 40%, 60%, and 80% quantiles to divide the data into 1-5) or logarithmic methods to achieve data smoothing and reduce the direct impact of extreme values ​​on sentencing results.

[0043] Step 2-1: Based on the preprocessed decision elements, build a multi-input single-output multi-layer neural network model, where the number of input layer nodes is determined by the number of decision elements, and the number of hidden layer nodes and the number of hidden layers are determined by the combined size of the data and the number of decision features.

[0044] Step 2-2: Divide the preprocessed structured data into training, validation, and test sets according to a certain ratio. Randomly shuffle the data in the training set. For the t-th case input into the neural network, calculate the initial fixed-term imprisonment sentence (in months) for that case in the feedforward neural network model using the following formula:

[0045]

[0046] Among them, X t W represents the data vector of judgment elements for the t-th case. i ,i = 1, 2 are the weight matrices between the input layer and the hidden layer, and between the hidden layer and the output layer, respectively; B i ,i=1,2 are the threshold vectors between the input layer and the hidden layer, and between the hidden layer and the output layer, respectively; and Let represent the estimated value of the above parameters when the t-th case is input into the saturated neural network, and max{x,0} represent the vector formed by taking the maximum value of each dimension of vector x after comparing it with 0.

[0047] Steps 2-3: Based on the common attribute characteristics of the sentencing ranges for specific crimes, a saturation constraint with floating upper and lower bounds is imposed on the nested output layers of the constructed neural network. The specific definition is as follows:

[0048]

[0049] in, The term of imprisonment obtained in step 2-2; The final declared sentence is obtained by adding saturation limits; U t It is the upper limit of the corresponding statutory penalty range, L t These are the lower limits of the corresponding statutory sentencing ranges, which vary depending on the nature of the case and fluctuate according to the nature of the statutory sentencing circumstances.

[0050] Furthermore, a schematic diagram of the saturated neural network sentencing model is shown below. Figure 2 As shown, the model constructs a feedforward neural network with a specific number of nodes based on the number of judgment elements and the amount of data for a specific crime, and nests a saturation constraint function with judicial interpretability in the output layer.

[0051] Step 3-1: To enable the estimated sentence to converge to the actual sentence more quickly, this disclosure proposes to unify the L1 modulus relative error formula to be optimized and the precision formula for measuring sentencing accuracy, and to calculate the average relative error and average relative precision on the network using the following formulas respectively:

[0052]

[0053]

[0054] Among them, z t Let t be the actual sentence value for the t-th case. This is the estimated sentence for the t-th case calculated by the saturated neural network model. The estimated sentence for the t-th case calculated by the neural network model before adding saturation constraints.

[0055] Step 3-2: Based on the constructed saturated neural network structure, using the backpropagation algorithm based on saturation constraints and average relative error (SC-MRE-BP), and taking the L1 relative error formula from Step 3-1 as the network optimization objective, calculate the gradient term using the following formula:

[0056]

[0057]

[0058] Where sgn(x) is the sign function; I A It is an indicator function; These are the gradient values ​​for the saturation region, which can be set by the user, but generally should not exceed 0.5; the operator ⊙ represents the Hadamard product of the matrix, that is, the multiplication of corresponding elements of the matrix; and i = 1, 2, representing the estimated values ​​of the corresponding parameters when the t-th case is input into the saturated neural network. For matrix The element in the i-th row and j-th column.

[0059] Subsequently, the step size update formulas for each parameter are derived using the gradient terms mentioned above:

[0060]

[0061]

[0062]

[0063]

[0064] in, i = 1, 2 are obtained from step 3-2, X t This represents the data vector of judgment elements for the t-th case. and i = 1, 2, representing the estimated values ​​of the corresponding parameters when the t-th case is input into the saturated neural network.

[0065] Step 3-3: Next, we use the Adam algorithm, which can adaptively adjust the learning rate, to update the parameter estimates. The specific parameter update formula is as follows:

[0066]

[0067] m t =β1m t-1 +(1-β1)Δθ t ,

[0068] v t =β2v t-1 +(1-β2)Δθ t ⊙Δθ t ,

[0069] Where, θ t This represents the parameter estimate when the t-th sample is trained, and can be arbitrarily replaced with the weight estimate. i = 1, 2 or threshold estimate i = 1, 2; learning rate η = 0.001; exponential decay rate β1 = 0.9, β2 = 0.99, smoothing coefficient ε = e -7 Both m0 and v0 are 0; Operator and This indicates that the corresponding operation is performed on each element of matrix x; e is a matrix of all ones in the corresponding dimension.

[0070] Furthermore, the specific process of the SC-MRE-BP algorithm is summarized as follows: Figure 3 As shown:

[0071] The input is a structured case training set D = {(X t , z t The learning rate, exponentially decaying learning rate, and smoothing coefficient are involved. The specific process includes:

[0072] 1: Randomly initialize all weights and thresholds within the range (0,1);

[0073] 2: Repeat the data in the training set, if all (X) t , z t If the case belongs to D, then calculate the estimated sentence for the current case according to the formulas in steps 2-2 to 2-3, calculate the gradient term and parameter update step size according to the formula in step 3-2, and then update the weights and thresholds according to the Adam algorithm in step 3-3.

[0074] 3: A saturated neural network with output weights and thresholds determined until the average relative error reaches the stopping condition MRE < 0.0065.

[0075] As an example, the saturated neural network model and SC-MRE-BP algorithm proposed in this disclosure can be used to perform sentencing calculations and analyses on different types of cases. Taking single-defendant cases of typical crimes such as intentional injury, fundraising fraud, counterfeiting registered trademarks, and selling counterfeit registered trademarks as examples, sentencing calculations were performed respectively, and the sentencing results were compared with the calculation accuracy of traditional neural network and linear regression models. The data details of each crime and the specific sentencing calculation results under some circumstances (relatively serious circumstances) are shown in Table 1 below:

[0076] Table 1 compares the accuracy of different sentencing methods for different cases.

[0077]

[0078] The definition of sentencing accuracy is given in step 3-1. It can be seen that, with only a small sample size of two to three thousand data points, the average sentencing accuracy of this disclosure can reach over 85%; for the calculation of sentences for each crime, the results of this disclosure are consistently better than the neural network model (average improvement of about 1.01%) and significantly better than the linear regression model (average improvement of about 8.56%).

[0079] Example 2

[0080] One embodiment of this disclosure provides a case-assisted analysis system based on a saturated neural network, comprising:

[0081] The data acquisition module is used to acquire case description text and judgment document information, and to perform text preprocessing, automatically extract feature fields, and perform special element smoothing after acquiring feature fields to form structured nonlinear data.

[0082] The case analysis quantification module is used to input structured nonlinear data into a saturated neural network model to assist in calculating the estimated sentencing time for a case. Combining the common attribute characteristics of the sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain auxiliary case analysis results.

[0083] Example 3

[0084] One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the case-assisted analysis method based on a saturated neural network.

[0085] Example 4

[0086] One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the case-assisted analysis method based on saturated neural networks.

[0087] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A case-aided analysis method based on saturated neural networks, characterized in that, include: The system acquires case description text and judgment documents, performs text preprocessing, automatically extracts feature fields, and performs special element smoothing processing after acquiring feature fields to form structured nonlinear data. Structured nonlinear data is input into a saturated neural network model to assist in calculating the estimated sentencing time for a case. Combining the common attribute characteristics of the sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain the auxiliary case analysis results. Among them, based on the decision elements obtained from preprocessing, a multi-input single-output saturated neural network model is built. The number of nodes in the input layer is determined by the number of decision elements, and the number of nodes and the number of hidden layers are determined by the combined size of the data and the number of decision features. The preprocessed structured data is divided into training, validation, and test sets according to a set ratio. The training set data is then randomly shuffled. For the first... For each case input into the neural network, the initial prison sentence is calculated in the feedforward neural network model: in, Indicates the first t Data vector of judgment elements for each case, These are the weight matrices between the input layer and the hidden layer, and between the hidden layer and the output layer, respectively. These are the threshold vectors between the input layer and the hidden layer, and between the hidden layer and the output layer, respectively; and Indicates when the first t The estimated values ​​of the corresponding parameters when a case is input into a saturated neural network. Represents a vector x The vector formed by taking the maximum value of each dimension after comparing it with 0; Based on the common attribute characteristics of the sentencing ranges for specific crimes, a saturation constraint with floating upper and lower bounds is established for the nested output layers of the neural network, which is specifically defined as follows: in, The estimated sentence for the t-th case calculated by the neural network model before adding saturation constraints; The final declared sentence is obtained by adding saturation limits; It is the upper limit of the corresponding statutory penalty range. It is the lower limit of the corresponding statutory sentencing range, which varies depending on the nature of the case and fluctuates according to the nature of the statutory sentencing circumstances; The first calculation for the saturated neural network model t Estimated sentences for each case; The average relative error and average relative accuracy on the network are calculated as follows: in, For the first t The actual sentence value for each case, The first calculation for the saturated neural network model t Estimated sentences for each case; The estimated sentence for the t-th case calculated by the neural network model before adding saturation constraints; Based on a saturated neural network model, and using a backpropagation algorithm based on saturation constraints and average relative error, L 1. The relative error formula is the optimization objective of the network, and the gradient term is calculated as follows: in, sgn ( x ) is a symbolic function; I {A} It is an indicator function; These are the gradient values ​​for the saturation region; operators Hadamard product represents the product of corresponding elements of a matrix; For matrix No. i Line 1 j Column elements; Subsequently, the step size update formulas for each parameter are derived further using the gradient term: in, Indicates the first t Data vector of judgment elements for each case, and Indicates when the first t The estimated values ​​of the corresponding parameters when a case is input into a saturated neural network; The parameter estimates are updated using the Adam algorithm, which can adaptively adjust the learning rate. The specific parameter update formula is as follows: in, Indicates the first t The parameter estimates of each sample during training can be arbitrarily replaced with weight estimates. or threshold estimate Learning rate Take 0.001; exponential decay rate Take 0.9, Set the smoothing coefficient to 0.

99. Pick ; and All values ​​are 0; Operators and Represents a matrix Perform corresponding operations on each element; It is a matrix of all 1s in the corresponding dimension.

2. The case-assisted analysis method based on a saturated neural network as described in claim 1, characterized in that, Preprocessing is performed using large language models and regularization methods. For elements with complex semantics and for which accurate text recognition rules cannot be established, Baidu Wenxin Yiyan is used to achieve fully automatic and accurate extraction. For elements with simple semantics and clear text recognition rules, regular expressions are used to achieve efficient extraction.

3. The case-assisted analysis method based on a saturated neural network as described in claim 2, characterized in that, Based on the distribution of specific features, some excessively sparse features are deleted to eliminate the influence of sparse features. Cases containing sparse features are also deleted accordingly, and finally, the main feature factors are determined in the modeling.

4. The case-assisted analysis method based on a saturated neural network as described in claim 1, characterized in that, The saturated neural network model is a multi-input single-output model structure. The number of input layer nodes is determined by the number of decision elements, and the number of hidden layer nodes and the number of hidden layers are determined by the combined effect of the data volume and the number of features.

5. A case-aided analysis system based on saturated neural networks, characterized in that, Specifically, the case-assisted analysis method based on saturated neural networks as described in any one of claims 1-4 includes: The data acquisition module is used to acquire case description text and judgment document information, and to perform text preprocessing, automatically extract feature fields, and perform special element smoothing after acquiring feature fields to form structured nonlinear data. The case analysis quantification module is used to input structured nonlinear data into a saturated neural network model to assist in calculating the estimated sentencing time for a case. Combining the common attribute characteristics of the sentencing intervals for specific crimes, a saturation constraint with floating upper and lower bounds is nested in the output layer of the saturated neural network model. At the same time, based on the saturation constraint and the backpropagation algorithm of the average relative error, with the relative error as the optimization objective of the network, the step size update of each parameter of the network is derived using the gradient term, and the parameter estimates are adaptively updated to obtain auxiliary case analysis results.

6. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the case-assisted analysis method based on a saturated neural network as described in any one of claims 1-4.

7. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform the case-assisted analysis method based on a saturated neural network as described in any one of claims 1-4.

Images