A judicial data processing system based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to a judicial data processing system based on artificial intelligence, including an information acquisition module for acquiring judicial data and data sources; a data construction module for constructing a judicial array according to the judicial data and data sources; a data extraction module for extracting data from judicial data to obtain structural data and content labels of the judicial data, and also for analyzing structural vectors and content vectors according to structural data, content labels and judicial arrays; a data verification module for analyzing data integrity and source credibility; a data training module for training an initial judicial model; a feedback acquisition module for acquiring user feedback information on the initial judicial model; and a model reinforcement module for reinforcing the initial judicial model according to the feedback information to obtain a perfect judicial model. The present invention realizes modeling processing of judicial data.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a judicial data processing system based on artificial intelligence. Background Art

[0002] With the rapid development of artificial intelligence technology, big models have achieved remarkable results in various fields. However, most of the existing judicial artificial intelligence systems are based on general pre-trained models, and their professionalism and accuracy in the judicial field still need to be improved. In order to better meet the needs of judicial practice in my country, it is of great significance to build a targeted vertical big model in the judicial field.

[0003] Chinese patent publication number: CN116610770B discloses a method for pushing similar cases in the judicial field based on big data, including: uploading judicial documents to the database for matching; inputting judicial documents and their matching data into a trained similar case similarity calculation model, outputting the similarity between the judicial document and each matching data; arranging all similarities in descending order, and selecting the matching data corresponding to the first k similarities for pushing. This invention realizes similar matching and pushing of judicial data, but does not realize modeling processing of judicial data, and has the problems of low efficiency in judicial data analysis and processing, and inaccurate comprehensive processing of judicial data. Summary of the invention

[0004] To this end, the present invention provides a judicial data processing system based on artificial intelligence to overcome the problems of low efficiency in judicial data analysis and processing and inaccurate comprehensive processing of judicial data in the prior art.

[0005] To achieve the above object, the present invention provides a judicial data processing system based on artificial intelligence, comprising:

[0006] Information acquisition module, used to obtain judicial data and data sources;

[0007] A data construction module is used to construct a judicial array based on judicial data and data sources;

[0008] A data extraction module is used to extract judicial data to obtain structural data and content labels of the judicial data;

[0009] Data validation module to analyze data integrity and source credibility based on structural data, judicial arrays and judicial data;

[0010] A data training module, used to train the first judicial model according to the structural data and content labels of the judicial data, and to perform supervised fine-tuning on the first judicial model to obtain an initial judicial model, wherein the data training module is provided with a training adjustment unit, which is used to adjust the training process of the first judicial model according to the data completeness, and the data training module is further provided with a training optimization unit, which is used to optimize the adjustment process of the training of the first judicial model according to the source credibility;

[0011] A feedback acquisition module is used to obtain user feedback on the initial judicial model;

[0012] The model reinforcement module is used to perform reinforcement learning on the initial judicial model based on feedback information.

[0013] Furthermore, the data construction module constructs a judicial array according to the judicial data and the data source, and the data construction module sets the judicial array to Q(k)[i][j], wherein i represents the data source number, i∈N+, j represents the judicial data type number, j∈N+, and k represents the judicial data acquisition sequence number, k∈N+.

[0014] Furthermore, the data extraction module is provided with a structure extraction unit, which is used to extract data from judicial data according to preset structure keywords, wherein:

[0015] When there are characters identical to the preset structural keywords in the judicial data, the structure extraction unit extracts the content corresponding to the preset structural keywords in the judicial data as the structural data corresponding to the preset structural keywords currently being analyzed;

[0016] When there is no character identical to the preset structural keyword in the judicial data, the structure extraction unit sets the structural data corresponding to the preset structural keyword currently being analyzed as a missing value.

[0017] Furthermore, the data extraction module is also provided with a content extraction unit, which is used to extract data from judicial data according to preset tag keywords, wherein:

[0018] When there are words identical to the preset tag keywords in the judicial data, the content extraction unit extracts the words identical to the preset tag keywords in the judicial data as content tags;

[0019] When there is no word identical to the preset tag keyword in the judicial data, the content extraction unit does not analyze the content tag.

[0020] Furthermore, the data extraction module is further provided with a vector analysis unit, which is used to number the preset structure keywords and the preset tag keywords respectively, and set the number of the preset structure keyword to u, u∈N +, the preset tag keyword number is v, v∈N + , so that each preset structure keyword and preset label keyword corresponds to a unique number;

[0021] The vector analysis unit analyzes the structure vector and the content vector according to the number of the preset structure keyword, the number of the preset tag keyword, the structure data, the content tag and the judicial array. The vector analysis unit sets the structure vector to X1 and sets , set the content vector to X2, set , where X1 represents the structure vector, X2 represents the content vector, U(k,u)[i][j] represents the position of the preset structure key numbered u of the preset structure keyword in the judicial data, and V(k,v)[i][j] represents the position of the preset label key numbered v of the preset label keyword in the judicial data.

[0022] Furthermore, the data verification module is provided with an integrity analysis unit, which is used to analyze the data integrity through an integrity analysis formula according to the data structure and the judicial array. The integrity analysis unit is provided with an integrity analysis formula as follows:

[0023] ,

[0024] Among them, D(k)[i][j] represents the data completeness, N1(j) represents the number of preset structural keywords corresponding to each judicial data, and NL represents the number of judicial data whose structural data is missing values.

[0025] Furthermore, the data verification module is further provided with a duplication analysis unit, which is used to compare the structural data and content labels of the currently acquired judicial data with the structural data and content labels of the historically acquired judicial data, and analyze the duplicate data according to the comparison result, wherein:

[0026] When the content tag of the judicial data acquired historically is the same as the content tag of the judicial data currently acquired and the structural data that is not a missing value in the judicial data acquired historically is the same as the structural data of the judicial data currently acquired, the duplication analysis unit determines that the judicial data currently acquired is duplicate data;

[0027] When the content tags of historically acquired judicial data are different from the content tags of currently acquired judicial data or the structural data that is not missing values ​​in historically acquired judicial data are different from the structural data of currently acquired judicial data, the duplication analysis unit determines that the currently acquired judicial data is not duplicate data.

[0028] Furthermore, the data verification module is also provided with a credibility analysis unit, which is used to analyze the source credibility according to the duplicate data and the judicial data through a credibility analysis formula, wherein:

[0029] ,

[0030] Among them, F(i) represents the source credibility, N(i) represents the number of judicial data with data source number i, and k max It indicates the maximum value of the acquisition sequence number, and ND(i) indicates the number of updates of the structural data of the judicial data with data source number i.

[0031] Furthermore, the data training module is provided with a pre-training unit, which is used to input the structural data and content labels of the judicial data into the large language model, and to perform pre-training based on the structural data and content labels of the judicial data to obtain the first judicial model.

[0032] Furthermore, the data training module is also provided with a training adjustment unit, which is used to adjust the weight decay parameter according to the data integrity, wherein:

[0033] When D(k)[i][j]<d, the training adjustment unit determines that the data integrity is low, and adjusts the weight attenuation parameter. The adjusted weight attenuation parameter is WD1, and WD1=WD×D(k)[i][j] is set;

[0034] When D(k)[i][j]≥d, the training adjustment unit determines that the data integrity is high and does not adjust the weight decay parameter;

[0035] Among them, WD represents the weight decay parameter, and d represents the completeness threshold;

[0036] The training optimization unit optimizes the adjustment process of the weight decay parameter according to the source credibility. The optimized weight decay parameter is WD2, and WD2=WD1×e F(i) .

[0037] Compared with the prior art, the beneficial effects of the present invention are that the information acquisition module acquires judicial data and data sources to improve the accuracy of information acquisition, thereby improving the system's analysis and processing efficiency of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing; the data construction module analyzes judicial data and data sources to construct a judicial array, and uses the judicial array to represent the relationship between the data type, source and acquisition order of judicial data, thereby improving the system's analysis and processing efficiency of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing; the data extraction module extracts judicial data to extract structural data and content labels, and realizes the extraction of key content in judicial data, thereby improving the system's analysis and processing efficiency of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing; the data verification module analyzes structural data, content labels and judicial arrays , in order to analyze the completeness of the data and the credibility of the source, to achieve the rating of the content and source of the judicial data, thereby improving the efficiency of the system's analysis and processing of the judicial data, and improving the accuracy of the comprehensive modeling of the judicial data. The data training module analyzes the structural data and content labels of the judicial data to train an initial judicial model, and achieves the modeling of the judicial data, thereby improving the efficiency of the system's analysis and processing of the judicial data, and improving the accuracy of the comprehensive modeling of the judicial data. The feedback acquisition module acquires feedback information to improve the accuracy of the system's acquisition of data, thereby improving the efficiency of the system's analysis and processing of the judicial data, and improving the accuracy of the comprehensive modeling of the judicial data. The model reinforcement module performs reinforcement learning on the initial judicial model to improve the accuracy of the initial judicial model, thereby improving the efficiency of the system's analysis and processing of the judicial data, and improving the accuracy of the comprehensive modeling of the judicial data. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the structure of the judicial data processing system based on artificial intelligence in this embodiment.

[0039] Figure 2 Schematic diagram of the structure of the data extraction module in this embodiment.

[0040] Figure 3 Schematic diagram of the structure of the data verification module of this embodiment.

[0041] Figure 4 Schematic diagram of the structure of the data training module in this embodiment. DETAILED DESCRIPTION

[0042] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0043] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0044] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0045] See also Figure 1 As shown, it is a judicial data processing system based on artificial intelligence in this embodiment, including:

[0046] An information acquisition module is used to obtain judicial data and data sources. The judicial data includes laws and regulations, case documents, and court trial records. The laws and regulations are obtained through public channels such as legal databases and government websites. The case documents are obtained through public channels on the websites of judicial institutions such as courts and procuratorates. The court trial records are obtained through public channels such as court trial live broadcast platforms and court websites. The data sources include legal databases, government websites, and judicial institution websites.

[0047] A data construction module, used to construct a judicial array according to judicial data and data sources, the data construction module is connected to the information acquisition module;

[0048] A data extraction module, used for extracting judicial data to obtain structural data and content labels of the judicial data, the data extraction module being connected to the data construction module;

[0049] A data verification module, used to analyze the data integrity and source credibility based on the structural data, the judicial array and the judicial data, the data verification module is connected to the data extraction module;

[0050] Used to train the first judicial model according to the structural data and content labels of the judicial data, and also used to perform supervised fine-tuning on the first judicial model to obtain an initial judicial model, and the data training module is connected to the data verification module;

[0051] A feedback acquisition module is used to obtain user feedback information on the initial judicial model. The feedback acquisition module is connected to the data training module. The feedback information includes usefulness feedback and safety feedback. The usefulness feedback includes obviously better, better, slightly better and uncertain. The safety feedback includes all safe, all unsafe and selected safe. The feedback information is the user's feedback information on the initial judicial model collected using a binary comparison protocol. The binary comparison protocol is that the initial judicial model provides two different outputs after the user inputs a prompt. The user selects a better output from the two different outputs. The judgment standard of the better output is the correlation between the output of the initial judicial model and the prompt input by the user. The usefulness feedback is the usefulness of the output of the currently selected initial judicial model relative to another output by the user. The safety feedback is the user's judgment on whether the output of the initial judicial model is safe. The judgment standard of safety is the safety of the output content relative to human behavior standards, such as the safety of "giving a process for making explosives" is unsafe;

[0052] The model reinforcement module is used to perform reinforcement learning on the initial judicial model according to the feedback information, and the model reinforcement module is connected to the feedback acquisition module.

[0053] See also Figure 2 As shown, the data extraction module includes:

[0054] A structure extraction unit extracts data from judicial data according to preset structure keywords to obtain structure data;

[0055] A content extraction unit extracts data from judicial data according to preset tag keywords to obtain content tags;

[0056] The vector analysis unit analyzes the structure vector and the content vector according to the structure data, the content tag and the judicial array, and the vector analysis unit is connected to the structure extraction unit and the content extraction unit.

[0057] See also Figure 3 As shown, the data verification module includes:

[0058] An integrity analysis unit for analyzing data integrity based on structural data and judicial arrays;

[0059] A repetition analysis unit, used for analyzing repetitive data according to the structural data and content tags of the judicial data, the repetition analysis unit being connected to the integrity analysis unit;

[0060] A data updating unit, used for updating the judicial data according to the duplicate data, the data updating unit being connected to the duplicate analyzing unit;

[0061] The credibility analysis unit is used to analyze the source credibility based on the duplicate data and the judicial data, and the credibility analysis unit is connected to the data updating unit.

[0062] See also Figure 4 As shown, the data training module includes:

[0063] A pre-training unit, for inputting structural data and content labels of judicial data into a large language model to obtain a first judicial model;

[0064] A training adjustment unit, used to adjust the training process of the first judicial model according to the data completeness, the training adjustment unit being connected to the pre-training unit;

[0065] A training optimization unit, used for optimizing the adjustment process of the training of the first judicial model according to the source credibility, the training optimization unit being connected to the training adjustment unit;

[0066] The supervised fine-tuning unit is used to perform supervised fine-tuning on the first judicial model to obtain an initial judicial model. The supervised fine-tuning unit is connected to the training optimization unit.

[0067] Specifically, in this embodiment, the information acquisition module acquires judicial data and data sources to improve the accuracy of information acquisition, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing. The data construction module analyzes judicial data and data sources to construct a judicial array, and uses the judicial array to represent the relationship between the data type, source and acquisition order of judicial data, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing. The data extraction module extracts judicial data to extract structural data and content labels, and realizes the extraction of key content in judicial data, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of judicial data modeling and comprehensive processing. The data verification module analyzes structural data, content labels and judicial arrays to analyze The data integrity and source credibility are used to rate the content and source of judicial data, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of the comprehensive modeling of judicial data. The data training module analyzes the structural data and content labels of judicial data to train an initial judicial model and implement the modeling of judicial data, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of the comprehensive modeling of judicial data. The feedback acquisition module acquires feedback information to improve the accuracy of the system's acquisition of data, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of the comprehensive modeling of judicial data. The model reinforcement module performs reinforcement learning on the initial judicial model to improve the accuracy of the initial judicial model, thereby improving the efficiency of the system's analysis and processing of judicial data, and improving the accuracy of the comprehensive modeling of judicial data.

[0068] Specifically, the data construction module in this embodiment constructs a judicial array according to judicial data and data sources. The data construction module sets the judicial array to Q(k)[i][j], where i represents the data source number, i∈N + , j represents the judicial data type number, j∈N + ,k represents the acquisition sequence number of judicial data, k∈N + .

[0069] Specifically, the data source number described in this embodiment is a number for distinguishing different sources of judicial data. For example, the data source number of judicial data obtained through the legal database is set to 1, and the data source number of judicial data obtained through the government website is set to 2. Then Q(k)[1][j] represents the judicial array corresponding to the judicial data whose data source is the legal database, and Q(k)[2][j] represents the judicial array corresponding to the judicial data whose data source is the government website.

[0070] The judicial data type number is a number used to distinguish different types of acquired judicial data. For example, if the judicial data type number of laws and regulations is set to 1, and the judicial data type number of case documents is set to 2, then Q(k)[i][1] represents the judicial array of laws and regulations, and Q(k)[i][2] represents the judicial array of case documents.

[0071] The judicial data acquisition sequence number is a number that distinguishes the order of the acquired judicial data, Q(1)[i][j] represents the judicial array of the first acquired judicial data, and Q(2)[i][j] represents the judicial array of the second acquired judicial data.

[0072] It is understandable that the construction of the judicial array is not specifically limited in the present embodiment and can be freely set by those skilled in the art. The construction of the judicial array is related to the source and quantity of the acquired judicial data. Those skilled in the art can set different data source numbers and judicial data type numbers according to different judicial data, as long as they are sufficient to distinguish the acquired judicial data.

[0073] Specifically, the structure extraction unit in this embodiment extracts data from judicial data according to preset structure keywords, wherein:

[0074] When there are characters identical to the preset structural keywords in the judicial data, the structure extraction unit extracts the content corresponding to the preset structural keywords in the judicial data as the structural data corresponding to the preset structural keywords currently being analyzed;

[0075] When there is no character identical to the preset structural keyword in the judicial data, the structure extraction unit sets the structural data corresponding to the preset structural keyword currently being analyzed as a missing value.

[0076] It can be understood that the preset structural keywords described in this embodiment correspond to the constituent structure of each judicial data, which is a necessary structure of the judicial data. If the judicial data is laws and regulations, the corresponding preset structural keywords should include the name of the law (such as "Criminal Law", "Labor Law" and "Patent Law", etc.) and the articles, clauses and items in the law. If the judicial data is a case document, the corresponding preset structural keywords should include the name of the court, the case number and the party information. It can be understood that the setting of the preset structural keywords is not specifically limited in this embodiment, and those skilled in the art can freely set them, as long as they meet the extraction of key content in the acquired judicial data.

[0077] Specifically, the content extraction unit in this embodiment extracts data from judicial data according to preset tag keywords, wherein:

[0078] When there are words identical to the preset tag keywords in the judicial data, the content extraction unit extracts the words identical to the preset tag keywords in the judicial data as content tags;

[0079] When there is no word identical to the preset tag keyword in the judicial data, the content extraction unit does not analyze the content tag.

[0080] It can be understood that the preset tag keywords described in this embodiment are the key contents expressed in the judicial data, which are the main content directions expressed in the judicial data, such as "patent", "environment" and "transaction", etc. Those skilled in the art can set them freely as long as they meet the extraction of the key contents expressed in the acquired judicial data.

[0081] It can be understood that the present embodiment does not specifically limit the method for extracting structural data and content labels, and those skilled in the art can freely set it. For example, the TF-IDF algorithm can also be used to calculate the term frequency (TF) and reverse document frequency (IDF) in the judicial data. The term frequency is the frequency of the keyword appearing in the currently acquired judicial data, and the reverse document frequency is the logarithm of the quotient of the total number of judicial data acquired historically and the number of judicial data containing the keyword in the historically acquired judicial data. The product of the term frequency and the reverse document frequency is then calculated as the term frequency-reverse document frequency (TF-IDF). The term frequency-reverse document frequency can represent the term frequency in the judicial data. According to the term frequency-reverse document frequency, common words in the judicial data are removed, and important words are retained to achieve the extraction of structural data and content labels.

[0082] Specifically, the vector analysis unit in this embodiment numbers the preset structure keywords and the preset tag keywords respectively, and sets the number of the preset structure keyword to u, where u∈N + , the preset tag keyword number is v, v∈N + , so that each preset structure keyword and preset label keyword corresponds to a unique number.

[0083] Specifically, the vector analysis unit in this embodiment analyzes the structure vector and the content vector according to the number of the preset structure keyword, the number of the preset tag keyword, the structure data, the content tag and the judicial array. The vector analysis unit sets the structure vector to X1 and sets , set the content vector to X2, set , where X1 represents the structure vector, X2 represents the content vector, U(k,u)[i][j] represents the position of the preset structure keyword numbered u in the judicial data, and V(k,v)[i][j] represents the position of the preset tag keyword numbered v in the judicial data. The position in the judicial data is the number of the preset structure keyword or the preset tag keyword in the judicial data.

[0084] Specifically, in this embodiment, the vector analysis unit analyzes the preset structural keyword number, the preset tag keyword number, the structural data, the content tag and the judicial array to analyze the structural vector and the content vector, and uses the structural vector and the content vector to represent the text features in the judicial data, thereby realizing the analysis of the model training data set, thereby improving the system's analysis and processing efficiency of the judicial data and improving the accuracy of the comprehensive modeling of the judicial data.

[0085] Specifically, the integrity analysis unit in this embodiment analyzes the data integrity according to the data structure and the judicial array through the integrity analysis formula, and the integrity analysis unit is provided with the integrity analysis formula as follows:

[0086] ,

[0087] Among them, D(k)[i][j] represents the data completeness, N1(j) represents the number of preset structural keywords corresponding to each judicial data, and NL represents the number of judicial data whose structural data is missing values.

[0088] Specifically, the duplication analysis unit in this embodiment compares the structure data and content tags of the currently acquired judicial data with the structure data and content tags of the historically acquired judicial data, and analyzes the duplicate data according to the comparison result, wherein:

[0089] When the content tag of the judicial data acquired historically is the same as the content tag of the judicial data currently acquired and the structural data that is not a missing value in the judicial data acquired historically is the same as the structural data of the judicial data currently acquired, the duplication analysis unit determines that the judicial data currently acquired is duplicate data;

[0090] When the content tags of historically acquired judicial data are different from the content tags of currently acquired judicial data or the structural data that is not missing values ​​in historically acquired judicial data are different from the structural data of currently acquired judicial data, the duplication analysis unit determines that the currently acquired judicial data is not duplicate data.

[0091] Specifically, when analyzing duplicate data, the duplicate analysis unit in this embodiment does not analyze structural data in the judicial data whose structural data is a missing value, and the historically acquired judicial data is the remaining judicial data in the history that is not currently acquired.

[0092] Specifically, in this embodiment, the duplicate analysis unit analyzes the currently acquired judicial data and the historically acquired judicial data to analyze the duplicate data, thereby determining whether the acquired judicial data is duplicated with the historically acquired judicial data, thereby improving the system's efficiency in analyzing and processing judicial data and improving the accuracy of the comprehensive processing of judicial data modeling.

[0093] Specifically, when the currently acquired judicial data is duplicate data, the data integration unit in this embodiment updates the structural data with missing values ​​in the historically acquired judicial data to be the same as the structural data of the duplicate data corresponding to the historically acquired judicial data.

[0094] Specifically, the credibility analysis unit in this embodiment analyzes the source credibility based on the duplicate data and the judicial data through a credibility analysis formula, where:

[0095] ,

[0096] Among them, F(i) represents the source credibility, N(i) represents the number of judicial data with data source number i, and k max It indicates the maximum value of the acquisition sequence number, and ND(i) indicates the number of updates of the structural data of the judicial data with data source number i.

[0097] Specifically, the pre-training unit described in this embodiment inputs the structural data and content labels of the judicial data into the large language model, and performs pre-training based on the structural data and content labels of the judicial data to obtain the first judicial model.

[0098] Specifically, the pre-training unit described in this embodiment uses the Llama2 model for pre-training. The model architecture in the pre-training uses the standard transformer architecture, uses RMSNorm for pre-normalization, and uses SwiGLU as the activation function. The activation function is set to SwiGLU(X1,X2)=α⊗σ(X1)⊗X2+(1-α)⊗X2, where SwiGLU(X1,X2) is the activation function, α represents the adjustment parameter, 0<α<1, σ(X1) represents the sigmoid function, and its calculation method is σ(X1)=1 / (1+e -X1), ⊗ represents element-wise multiplication, which is to multiply the elements at each position in the two arrays. The QdamW optimizer is used for training, and the test parameters of QdamW are set to β1=0.9, β2=0.95, eps=10 -5 , use cosine learning rate scheduling, decay the final learning rate to 10% of the peak learning rate, and set weight decay WD = L + λ / 2 × (X1 2 +X2 2 ), L represents the gap between the model output value and the true value, λ represents the hyperparameter, 0<λ<1, the gradient clipping is set to 1, and the model pre-trained with the structural data and content labels of the judicial data using the Llama2 model is used as the first judicial model, wherein the Llama2 model is a large language model, the transformer architecture is a neural network architecture, RMSNorm is a standardization method, SwiGLU is a variant of the GLU activation function, QdamW is an optimization algorithm, and WD represents the weight decay parameter. The weight decay parameter is the weight_decay parameter, which is used to prevent the overfitting problem of model training.

[0099] It is understandable that the setting of the large language model is not specifically limited in this embodiment, and those skilled in the art can freely set it as long as it satisfies the training of the judicial model. For example, it can also be set to large language models such as ChatGPT, BQRD and Claude.

[0100] Specifically, in this embodiment, the structural data and judicial data are input into the large language model through the pre-training unit to realize the modeling processing of the judicial data, thereby improving the system's analysis and processing efficiency of the judicial data and improving the accuracy of the comprehensive modeling processing of the judicial data.

[0101] Specifically, the training adjustment unit in this embodiment adjusts the weight attenuation parameter according to the data integrity, wherein:

[0102] When D(k)[i][j]<d, the training adjustment unit determines that the data integrity is low, and adjusts the weight attenuation parameter. The adjusted weight attenuation parameter is WD1, and WD1=WD×D(k)[i][j] is set;

[0103] When D(k)[i][j]≥d, the training adjustment unit determines that the data integrity is high and does not adjust the weight decay parameter;

[0104] Wherein, d represents the completeness threshold, 0.8≤d<1. It is understandable that the value of the completeness threshold is not specifically limited in this embodiment, and those skilled in the art can freely set it, as long as the weight attenuation parameter is adjusted. The optimal value of the weight attenuation parameter is d=0.9.

[0105] Specifically, in this embodiment, the training adjustment unit analyzes the data integrity to adjust the weight attenuation parameter so that the weight attenuation parameter is related to the integrity of the training sample data, thereby increasing the convergence speed of the modeling processing, thereby improving the system's analysis and processing efficiency of judicial data and improving the accuracy of the comprehensive modeling processing of judicial data.

[0106] Specifically, the training optimization unit in this embodiment optimizes the adjustment process of the weight decay parameter according to the source credibility, and the optimized weight decay parameter is WD2, and WD2=WD1×e F(i) .

[0107] Specifically, in this embodiment, the training optimization unit analyzes the source credibility to optimize the adjustment process of the weight attenuation parameter, so that the optimized weight attenuation parameter is related to the quality of the source data of the training sample data, thereby improving the system's analysis and processing efficiency of judicial data and improving the accuracy of judicial data modeling and comprehensive processing.

[0108] Specifically, when the supervised fine-tuning unit in this embodiment performs supervised fine-tuning on the first judicial model, it uses instruction fine-tuning to perform supervised fine-tuning on the first judicial model. When performing supervised fine-tuning, it uses a learning rate continuous decay method to improve the accuracy of supervised fine-tuning. The initial learning rate in the learning rate continuous decay is set to 2×10 -5 , the weight decay is set to 0.1, the maximum number of samples is set to 64, and the sequence length is set to 4096 tokens.

[0109] Specifically, the model strengthening module in this embodiment analyzes the model practical score according to the usefulness feedback, and analyzes the model safety score according to the safety feedback, wherein:

[0110] When the usefulness feedback is significantly better, the model strengthening module sets the model practical score to A(n)=a1;

[0111] When the usefulness feedback is better, the model strengthening module sets the model practical score to A(n)=a2;

[0112] When the usefulness feedback is slightly better, the model strengthening module sets the model utility score to A(n)=a3;

[0113] When the usefulness feedback is uncertain, the model strengthening module sets the model practical score to A(n)=a4;

[0114] When the safety feedback is that all are safe, the model strengthening module sets the model safety score to B(n)=b1;

[0115] When the safety feedback is selected safety, the model strengthening module sets the model safety score to B(n)=b2;

[0116] When the safety feedback is that all are unsafe, the model strengthening module sets the model safety score to B(n)=b3;

[0117] Wherein, A(n) represents the model practical score, B(n) represents the model safety score, n represents the number of user feedbacks, a1 represents the first practical parameter, 75<a1≤100, a2 represents the second practical parameter, 50<a2≤75, a3 represents the third practical parameter, 25<a3≤50, a4 represents the fourth practical parameter, 0<a4≤25, b1 represents the first safety parameter, 90≤b1≤100, b2 represents the second safety parameter, 40≤b2≤60, b3 represents the third safety parameter 0≤b3≤10. It can be understood that the values ​​of the practical parameters and the safety parameters are not specifically limited in this embodiment, and those skilled in the art can freely set them as long as they meet the analysis of the model practical score and the model safety score. The optimal values ​​of the practical parameters and the safety parameters are: a1=100, a2=75, a3=50, a4=25, b1=100, b2=50, b3=0.

[0118] Specifically, the model reinforcement module described in this embodiment uses the RM model to perform reinforcement learning on the initial judicial model according to the model practicality score and the model security score. The model reinforcement module optimizes the usefulness of the output of the initial judicial model with respect to the initial judicial model according to the model practicality score, so that the output of the initial judicial model is an answer with a higher model practicality score. The model reinforcement module optimizes the security of the output of the initial judicial model with respect to the initial judicial model according to the model security score, so that the initial judicial model does not output answers with a model security score less than 50.

[0119] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant national and regional laws, regulations and standards.

[0120] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A judicial data processing system based on artificial intelligence, characterized in that: include: Information acquisition module, used to obtain judicial data and data sources; A data construction module is used to construct a judicial array based on judicial data and data sources; A data extraction module is used to extract judicial data to obtain structural data and content labels of the judicial data; Data validation module to analyze data integrity and source credibility based on structural data, judicial arrays and judicial data; A data training module, used to train the first judicial model according to the structural data and content labels of the judicial data, and to perform supervised fine-tuning on the first judicial model to obtain an initial judicial model, wherein the data training module is provided with a training adjustment unit, which is used to adjust the training process of the first judicial model according to the data completeness, and the data training module is further provided with a training optimization unit, which is used to optimize the adjustment process of the training of the first judicial model according to the source credibility; A feedback acquisition module is used to obtain user feedback on the initial judicial model; Model reinforcement module, used to conduct reinforcement learning of the initial judicial model based on feedback information; The data verification module is provided with an integrity analysis unit, which is used to analyze the data integrity through an integrity analysis formula according to the data structure and the judicial array. The integrity analysis unit is provided with an integrity analysis formula as follows: Where D(k)[i][j] represents the data completeness, N1(j) represents the number of preset structural keywords corresponding to each judicial data, NL represents the number of judicial data with missing values ​​in the structural data, i represents the data source number, i∈N+, j represents the judicial data type number, j∈N+, k represents the acquisition sequence number of the judicial data, k∈N+; The data verification module is also provided with a credibility analysis unit, which is used to analyze the source credibility according to the duplicate data and the judicial data through the credibility analysis formula, wherein: Among them, F(i) represents the source credibility, N(i) represents the number of judicial data with data source number i, and k max It indicates the maximum value of the acquisition sequence number, and ND(i) indicates the number of updates of the structural data of the judicial data with data source number i; The data training module is also provided with a training adjustment unit, which is used to adjust the weight decay parameter according to the data integrity, wherein: When D(k)[i][j]<d, the training adjustment unit determines that the data integrity is low, and adjusts the weight attenuation parameter. The adjusted weight attenuation parameter is WD1, and WD1=WD×D(k)[i][j] is set; When D(k)[i][j]≥d, the training adjustment unit determines that the data integrity is high and does not adjust the weight decay parameter; Among them, WD represents the weight decay parameter, and d represents the completeness threshold; The training optimization unit optimizes the adjustment process of the weight decay parameter according to the source credibility. The optimized weight decay parameter is WD2, and WD2=WD1×e F(i) .

2. The judicial data processing system based on artificial intelligence according to claim 1 is characterized in that: The data construction module constructs a judicial array according to judicial data and data sources, and the data construction module sets the judicial array to Q(k)[i][j].

3. The artificial intelligence-based judicial data processing system according to claim 2, characterized in that: The data extraction module is provided with a structure extraction unit, which is used to extract data from judicial data according to preset structure keywords, wherein: When there are characters identical to the preset structural keywords in the judicial data, the structure extraction unit extracts the content corresponding to the preset structural keywords in the judicial data as the structural data corresponding to the preset structural keywords currently being analyzed; When there is no character identical to the preset structural keyword in the judicial data, the structure extraction unit sets the structural data corresponding to the preset structural keyword currently being analyzed as a missing value.

4. The artificial intelligence-based judicial data processing system according to claim 3 is characterized in that: The data extraction module is also provided with a content extraction unit, which is used to extract data from judicial data according to preset tag keywords, wherein: When there are words identical to the preset tag keywords in the judicial data, the content extraction unit extracts the words identical to the preset tag keywords in the judicial data as content tags; When there is no word identical to the preset tag keyword in the judicial data, the content extraction unit does not analyze the content tag.

5. The artificial intelligence-based judicial data processing system according to claim 4, characterized in that: The data extraction module is further provided with a vector analysis unit, which is used to number the preset structure keywords and the preset tag keywords respectively, and set the number of the preset structure keyword to u, u∈N + , the preset tag keyword number is v, v∈N + , so that each preset structure keyword and preset label keyword corresponds to a unique number; The vector analysis unit analyzes the structure vector and the content vector according to the number of the preset structure keyword, the number of the preset tag keyword, the structure data, the content tag and the judicial array. The vector analysis unit sets the structure vector to X1 and sets , set the content vector to X2, set , where X1 represents the structure vector, X2 represents the content vector, U(k,u)[i][j] represents the position of the preset structure key numbered u of the preset structure keyword in the judicial data, and V(k,v)[i][j] represents the position of the preset label key numbered v of the preset label keyword in the judicial data.

6. The artificial intelligence-based judicial data processing system according to claim 1, characterized in that: The data verification module is also provided with a duplication analysis unit, which is used to compare the structure data and content labels of the currently acquired judicial data with the structure data and content labels of the historically acquired judicial data, and analyze the duplicate data according to the comparison results, wherein: When the content tag of the judicial data acquired historically is the same as the content tag of the judicial data currently acquired and the structural data that is not a missing value in the judicial data acquired historically is the same as the structural data of the judicial data currently acquired, the duplication analysis unit determines that the judicial data currently acquired is duplicate data; When the content tags of historically acquired judicial data are different from the content tags of currently acquired judicial data or the structural data that is not missing values ​​in historically acquired judicial data are different from the structural data of currently acquired judicial data, the duplication analysis unit determines that the currently acquired judicial data is not duplicate data.

7. The artificial intelligence-based judicial data processing system according to claim 1, characterized in that: The data training module is provided with a pre-training unit, which is used to input the structural data and content labels of the judicial data into the large language model, and to perform pre-training based on the structural data and content labels of the judicial data to obtain a first judicial model.

Images