Judgment prediction method and system for judicial case, and medium

Abstract

The invention discloses a judicial case judgment and prediction method and system and a medium. Relates to the technical field of natural language processing. According to the scheme, the method is improved on the basis of the traditional prediction technology, the fact description feature representation model for extracting the fact description part is obtained based on the judicial judgment document and the content, the content and structural integrity of the judicial judgment document is guaranteed to the maximum extent, the text length is reduced, and the prediction efficiency is improved. On the other hand, a fact description feature representation model fitting a real case judgment scene is constructed based on an xsBert model and a BiGRU model, and the fact description feature representation is mapped to criminal names and law articles by a full connection layer, so that the judgment prediction problem of a plurality of criminal names and law articles is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method, system and medium for predicting the judgment of a judicial case. Background Art

[0002] In today's era, natural language processing (NLP) technology is experiencing rapid development and is widely used in text analysis, machine translation, chatbots and other fields. With the rise of deep learning, models such as BERT and GPT have become standards, capable of understanding complex language structures and contexts, significantly improving the ability to generate and understand text. At the same time, pre-training and fine-tuning strategies allow these models to quickly adapt to various specific tasks, lowering the development threshold. In addition, emerging research directions such as cross-language processing and multimodal learning have further expanded the application scenarios of natural language processing technology and promoted the advancement of technologies such as intelligent voice assistants and content recommendation systems.

[0003] In the judicial field, natural language processing also plays an important role. First, natural language processing technology can automatically process a large amount of legal documents and data, significantly reducing the time judges and lawyers spend on document analysis, retrieval and organization, and improving case handling efficiency. Secondly, through intelligent algorithms, natural language processing technology can ensure more accurate analysis of legal clauses and cases, reduce the risk of human errors and inconsistencies, and improve judicial fairness. In addition, natural language processing technology can transform complex legal language into easy-to-understand content, making it easier for the public to obtain legal information and enhancing the popularization and transparency of legal knowledge. In addition, natural language processing technology can analyze historical cases and judgments, provide data support and trend forecasts, and help judges and lawyers make more informed decisions in case trials and strategy formulation.

[0004] At present, most of the better models in the field of judicial judgment prediction are aimed at predicting single-charge judgments, and they are studied as a single-label multi-classification problem. When preprocessing the corpus, cases with multiple charges and legal provisions are directly removed, which makes the problem of one case involving multiple charges and legal provisions not sufficiently studied. However, in the process of actual trial of cases, some cases usually involve multiple charges, and may be convicted and sentenced according to multiple legal provisions, and may even involve multiple crimes. However, due to the complexity of cases with multiple charges and multiple legal provisions, there is not much research involved at present. In addition, the current neural networks such as CNN and RNN extract text features of case facts as input to predict the corresponding legal provisions and charge labels, which has corresponding drawbacks. RNN networks may face the problem of gradient vanishing when processing long texts, resulting in insufficient understanding of long-range dependencies; CNN networks usually have difficulty in capturing global context information and focus mainly on local features. Both may be affected by input noise, resulting in inaccurate feature extraction, thereby affecting the performance of downstream tasks. Judicial text belongs to the category of long texts. The above methods are not effective in extracting features from long texts, and are more sensitive to the influence of noise, resulting in inaccurate judgment results and other problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing judicial decision prediction methods lack research on cases involving multiple crimes and legal provisions, the feature extraction of judicial texts is not complete, and there is a problem of information loss; the purpose of the present invention is to provide a judicial case decision prediction method, system and medium, to improve the method on the basis of traditional prediction technology, to obtain a fact description feature representation model for extracting the fact description part based on the judicial decision documents and contents, to ensure the content and structural integrity of the judicial decision documents to the greatest extent, to reduce the text length, to avoid the problem of incomplete original text loading caused by the input length limitation of the prediction model, on the other hand, to construct a fact description feature representation model that fits the real case judgment scenario based on the xsBert model and the BiGRU model, to map the fact description feature representation to the crime and legal provisions by the fully connected layer, to solve the problem of judgment prediction involving multiple crimes and legal provisions.

[0006] The present invention is achieved through the following technical solutions:

[0007] This solution provides a method for predicting the judgment of a judicial case, including:

[0008] Obtaining original data of judicial judgment documents, and preprocessing the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part, and a legal article judgment part;

[0009] Based on the xsBert model, feature encoding is performed on the fact description part to obtain the semantic encoding information of the fact description, and bidirectional feature extraction is performed on the semantic encoding information of the fact description based on the BiGRU model to obtain the feature representation model of the fact description;

[0010] Construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and use the fully connected layer to map the fact description feature representation to the crime and legal provisions to obtain the probability value of the judicial case to be predicted on each crime and legal provision;

[0011] A threshold value is judged on the probability value to obtain a judgment prediction result.

[0012] A further optimization scheme is that the original data of the judicial judgment document is obtained, and the original data is preprocessed to obtain a judgment data set, including the following method:

[0013] Obtaining original data of judicial decision documents;

[0014] Based on the fact description keywords, the crime judgment keywords and the legal article judgment keywords, the fact description part, the crime judgment part and the legal article judgment part are extracted respectively;

[0015] Extract crime labels and legal article labels from the crime judgment part and the legal article judgment part to form a label list;

[0016] Set the specified crime label and the specified legal article label to filter out the original data that does not contain the specified crime label and the specified legal article label;

[0017] Construct a judgment dataset of <fact description content, crime label set, and legal provision label set>.

[0018] A further optimization scheme is that the feature encoding of the fact description part is performed based on the xsBert model to obtain the semantic encoding information of the fact description, and bidirectional feature extraction is performed on the semantic encoding information of the fact description based on the BiGRU model to obtain the feature representation model of the fact description, including the method:

[0019] Obtain the fact description content of the judgment data set, insert the [CLS] symbol at the first position of each fact description sentence in the fact description content, and then input it into the xsBert model for encoding;

[0020] Obtain the semantic feature vector of each fact description sentence based on the xsBert model;

[0021] The semantic feature vector is mapped to the same latent space based on the BiGRU network, the fine-grained features of the text are extracted, and a fact description feature representation model is obtained according to the fine-grained features of the text.

[0022] A further optimization scheme is that the network structure of the xsBert model includes an input layer, a hidden layer and an output layer;

[0023] The number of hidden layers is 12 in total, and the hidden layers include an attention layer, a fully connected layer, an intermediate layer, and a hidden output layer;

[0024] The input layer obtains the word vector of the fact description sentence as an input vector;

[0025] The attention layer is a multi-head attention mechanism with 12 heads; after the attention layer obtains the attention score of the fact description sentence, it superimposes the input vector as the output;

[0026] The output of the attention layer is connected to a fully connected layer, and then the output of the middle layer is obtained through the activation function GELU;

[0027] The hidden output layer performs full connection and Dropout operations on the output of the intermediate layer to obtain the output of the hidden output layer;

[0028] The output of all hidden output layers is used as the output of the xsBert model.

[0029] A further optimization scheme is that the semantic feature vector is mapped to the same hidden space based on the BiGRU network, the fine-grained features of the text are extracted, and a fact description feature representation model is obtained according to the fine-grained features of the text, including the following method:

[0030] The semantic feature vectors are fused to obtain a fusion vector;

[0031] Initialize a bidirectional GRU network and input the fusion vector into the bidirectional GRU network to capture the fine-grained features of the text;

[0032] Concatenate the output representations of the last two time steps in the bidirectional GRU network to obtain the final fact description feature representation model.

[0033] A further optimization scheme is that the text fine-grained feature output is expressed as:

[0034] output=GRU(fact_all,num_layer,Dim,bidirectional,dropout)

[0035] Among them, Dim is the model encoding dimension, num_layer is the number of network layers; fact_all represents the fusion vector; bidirectional indicates whether it is bidirectional encoding; dropout represents the probability of random discarding to prevent overfitting.

[0036] A further optimization scheme is to construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, map the fact description feature representation to the crime and legal provisions by the fully connected layer, and obtain the probability value of the judicial case to be predicted on each crime and legal provision; including the method:

[0037] Obtain the original data of the case to be predicted and construct a case description label dataset of <fact description content, crime label set, and legal article label set>;

[0038] The case description label data set is input into the fact description feature representation model, and the features are mapped to legal provisions and crimes through a fully connected layer to obtain the distribution probability of the case to be predicted in each legal provision label and crime label;

[0039] The distribution probability of the case to be predicted in each legal article label and crime label is input into the Sigmoid activation function to calculate the probability value of the judicial case to be predicted in each crime and legal article.

[0040] A further optimization scheme is that the threshold value judgment is performed on the probability value to obtain the judgment prediction result, including a method: the probability value of the judicial case to be predicted on each crime and legal provision is mapped into the real crime and legal provision through a hash table.

[0041] This solution provides a judgment prediction system for judicial cases, which is used to implement the judgment prediction method of a judicial case mentioned above, and the system includes:

[0042] A preprocessing module is used to obtain the original data of the judicial judgment document and preprocess the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part and a legal article judgment part;

[0043] A feature extraction module is used to perform feature encoding on the fact description part based on the xsBert model to obtain semantic encoding information of the fact description, and to perform bidirectional feature extraction on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description;

[0044] The prediction module is used to construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and map the fact description feature representation to the crime name and legal provision by the fully connected layer to obtain the probability value of the judicial case to be predicted on each crime name and legal provision;

[0045] The output module is used to perform threshold judgment on the probability value to obtain a judgment prediction result.

[0046] The present solution also provides a computer-readable medium on which a computer program is stored. The computer program is executed by a processor to implement the above-mentioned method for predicting the decision of a judicial case.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] 1. The present invention provides a judgment prediction method, system and medium for judicial cases. This solution improves the method on the basis of traditional prediction technology, obtains a fact description feature representation model for extracting the fact description part based on the judicial judgment document and content, ensures the content and structural integrity of the judicial judgment document to the greatest extent, reduces the text length, and avoids the problem of incomplete original text loading caused by the input length limitation of the prediction model. On the other hand, based on the xsBert model and the BiGRU model, a fact description feature representation model that fits the real case judgment scenario is constructed, and the fact description feature representation is mapped to the crime and legal provisions by the fully connected layer, so as to solve the judgment prediction problem involving multiple crimes and legal provisions.

[0049] 2. The present invention provides a judgment prediction method, system and medium for judicial cases; this solution performs regular matching extraction on the content of judicial judgment documents based on the document structure of the judicial judgment documents to obtain the fact description part and the final judgment part recognized by the court, shorten the text length, and improve the feature extraction effect; based on the xsBert model, the fact description part is subjected to feature extraction, and the pre-trained model is used to fully semantically encode the fact description part, making it more suitable for the judicial field. Based on the BiGRU model network, the semantic information feature vector is mapped to the same latent space to further capture the fine-grained feature information in the fact description. The linear layer is used as the prediction output, and the extracted features are output predicted based on the multi-label classification prediction method to obtain the final prediction result, making it more suitable for the actual judgment prediction scenario, including the scenario of multiple charges and laws.

[0050] 3. The present invention provides a judgment prediction method, system and medium for judicial cases; this solution performs regular matching on judicial judgment documents according to the document structure to extract the fact description part and the judgment prediction part, which can ensure the content and structural integrity of the judicial judgment documents to the greatest extent, shorten the text length, and avoid the problem of incomplete original text loading due to the model input length limitation.

[0051] 4. The present invention provides a method, system and medium for predicting judgments in judicial cases. This solution combines the xsBert model + BiGRU model to jointly solve the impact of model input or output limitations on long text feature extraction, and uses the xsBert model to extract features from single sentences. Since the length of a single sentence is often limited, it can effectively circumvent the model input length limitation. The BiGRU model is used for fine-grained feature extraction, which can make up for the defect that the xsBert model cannot capture the contextual semantics between single sentences due to extracting single sentence features. The linear layer is used for multi-label classification prediction, which makes the model more in line with the actual case judgment scenario, and can solve the judgment prediction problem involving multiple crimes and laws. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0053] Figure 1 A flowchart of a method for predicting judgments in judicial cases;

[0054] Figure 2 A schematic diagram of the judgment prediction process for judicial cases;

[0055] Figure 3 Schematic diagram of the judgment prediction system for judicial cases. DETAILED DESCRIPTION

[0056] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0057] The existing judicial decision prediction methods lack research on cases involving multiple crimes and laws, and the feature extraction of judicial texts is not complete, resulting in information loss. In view of the shortcomings of the existing technology, this solution provides the following embodiments to solve the above technical problems:

[0058] Example 1

[0059] This embodiment provides a method for predicting the judgment of a judicial case. Figure 1 and Figure 2 As shown, including:

[0060] Step 1: obtain the original data of the judicial judgment document, and pre-process the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part, and a legal judgment part; Step 1 specifically includes the following sub-steps:

[0061] S11, obtain the original data of judicial decision documents;

[0062] S12, based on the fact description keywords, the crime judgment keywords and the legal article judgment keywords, the fact description part, the crime judgment part and the legal article judgment part are extracted respectively; in the specific implementation process, the crime judgment keywords and the legal article judgment keywords include "the trial found...", "This court believes...", "The judgment is as follows...", "The ruling is as follows...", "According to the provisions of the law...", etc. in the judicial judgment documents.

[0063] S13, extracting crime labels and legal article labels from the crime judgment part and the legal article judgment part to form a label list;

[0064] S14, setting a designated crime label and a designated legal article label, and filtering out the original data that does not contain the designated crime label and the designated legal article label;

[0065] S15, construct a judgment dataset of <fact description content, crime label set, and legal provision label set>.

[0066] Identify and skip the basic information in judicial judgment documents, such as information related to personnel information, location information, etc. that does not involve the content of the judgment; read the content of the judicial judgment document sentence by sentence, identify the fact description part recognized by the court, the crime judgment part and the legal judgment part; if the corresponding content is not identified in the judicial document, the judgment document is abandoned and considered incomplete; continue to process the identified content, convert the crime label and legal label into corresponding digital representation, and form a label list for content containing multiple labels; filter out data that does not contain the specified label to form a real and complete judgment data set.

[0067] Step 2: Based on the xsBert model, feature encoding is performed on the fact description part to obtain semantic encoding information of the fact description, and bidirectional feature extraction is performed on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description; this step specifically includes the following sub-steps:

[0068] S21, obtain the fact description content of the judgment data set: Fact = {fact1, fact2, fact3, ...fact n}, where n represents the total number of fact description sentences in the fact description content; n Represents the nth fact description sentence.

[0069] The network structure of the xsBert model described in this scheme includes an input layer, a hidden layer and an output layer; the hidden layer has a total of 12 layers, and the hidden layer includes an attention layer, a fully connected layer, an intermediate layer and a hidden output layer; the input layer obtains the word vector of the fact description sentence as the input vector; the attention layer is a multi-head attention mechanism with 12 heads; after the attention layer obtains the attention score of the fact description sentence, it superimposes the input vector as the output; the output of the attention layer is connected to a fully connected layer, and then the output of the intermediate layer is obtained through the activation function GELU; the hidden output layer performs full connection and Dropout operations on the output of the intermediate layer to obtain the output of the hidden output layer; the output of all hidden output layers is used as the output of the xsBert model.

[0070] S22, obtaining the semantic feature vector of each fact description sentence based on the xsBert model;

[0071] According to the xsBert model, the maximum length of the input text is N, N = 380, for each fact description sentence fact i ={w1, w2, w3, ... w L}, where L is the length of the fact description sentence, L∈N; insert the [CLS] symbol before w1 of each fact description sentence, and replace the fact description sentence after the [CLS] symbol with the fact i，cls The input is encoded into the xsBert model, where a [CLS] flag is added before the corpus, and the vector corresponding to this bit in the last layer can be used as the semantic representation of the entire sentence, so as to be used for downstream classification tasks; the features of each fact description sentence are extracted through the xsBert model to obtain the output vector The fact i，cls It represents the output representation after the [CLS] symbol is inserted into the i-th sentence in the fact description;

[0072] S23, mapping the semantic feature vector to the same latent space based on the BiGRU network, extracting the fine-grained features of the text, and obtaining a fact description feature representation model based on the fine-grained features of the text. This step specifically includes the following method:

[0073] S231, fuse the semantic feature vectors to obtain the fused vector fact all:

[0074]

[0075] Where n represents the number of sentences in a fact description content;

[0076] S232, initialize a bidirectional GRU network, input the fusion vector into the bidirectional GRU network to capture the fine-grained features of the text; the fine-grained features of the text output are expressed as:

[0077] output=GRU(fact_all,num_layer,Dim,bidirectional,dropout)

[0078] Among them, Dim represents the model encoding dimension, num_layer represents the number of network layers; fact_all represents the fusion vector; bidirectional represents whether it is bidirectional encoding; dropout represents the probability of random discarding to prevent overfitting.

[0079] S233, concatenate the output representations of the last two time steps in the bidirectional GRU network to obtain the final fact description feature representation model f: f = Concate(output[-2,:,:], output[-1,:,:]);

[0080] Where output[-2, :, :] and output[-1, :, :] represent the outputs of the last two time steps of the GRU network.

[0081] Step 3: Construct a case description label dataset for the case to be predicted, input the case description label dataset into the fact description feature representation model, and map the fact description feature representation to the crime and legal provisions by the fully connected layer to obtain the probability value of the judicial case to be predicted on each crime and legal provision; this step specifically includes the following methods:

[0082] S31, obtaining the original data of the case to be predicted, and constructing a case description label data set of <fact description content, crime label set, and legal article label set>;

[0083] S32, input the case description label data set into the fact description feature representation model, map the features to the legal provisions and crimes through a fully connected layer, and obtain the distribution probability of the case to be predicted in each legal provision label and crime label; the distribution probability of the case to be predicted i is: R i =Linear i (f); where f represents the final fact description feature expression.

[0084] S33, input the distribution probability of the case to be predicted in each legal article label and crime label into the Sigmoid activation function to calculate the probability value of the judicial case to be predicted in each crime and legal article: R out，i =Sigmoid(R i );where R i Represents the distribution probability of case i to be predicted.

[0085] Step 4: Perform threshold judgment on the probability value to obtain a judgment prediction result.

[0086] This step specifically includes the following method: mapping the probability values ​​of the judicial case to be predicted on each crime and legal provision into the actual crime and legal provision through a hash table.

[0087] This solution extracts the content of judicial judgment documents based on the document structure of the judicial judgment documents by regular matching, obtains the fact description part recognized by the court and the final judgment part, shortens the text length, and improves the feature extraction effect; based on the xsBert model, the fact description part is extracted, and the pre-trained model can be used to fully semantically encode the fact description part to make it more suitable for the judicial field. Based on the BiGRU model network, the semantic information feature vector is mapped to the same latent space to further capture the fine-grained feature information in the fact description. The linear layer is used as the prediction output, and the extracted features are output predicted based on the multi-label classification prediction method to obtain the final prediction result, which is more suitable for the actual judgment prediction scenario, including scenarios with multiple charges and laws.

[0088] Example 2

[0089] This embodiment provides a judgment prediction system for judicial cases. Figure 3 As shown, a method for predicting a judgment of a judicial case described in Example 1 is implemented, and the system includes:

[0090] A preprocessing module is used to obtain the original data of the judicial judgment document and preprocess the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part and a legal article judgment part;

[0091] A feature extraction module is used to perform feature encoding on the fact description part based on the xsBert model to obtain semantic encoding information of the fact description, and to perform bidirectional feature extraction on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description;

[0092] The prediction module is used to construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and map the fact description feature representation to the crime name and legal provision by the fully connected layer to obtain the probability value of the judicial case to be predicted on each crime name and legal provision;

[0093] The output module is used to perform threshold judgment on the probability value to obtain a judgment prediction result.

[0094] Example 3

[0095] This embodiment provides a computer-readable medium on which a computer program is stored. The computer program is executed by a processor to implement a judgment prediction method for a judicial case described in Embodiment 1, which specifically includes the following steps:

[0096] Step 1: obtaining the original data of the judicial judgment document, and preprocessing the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part, and a legal judgment part;

[0097] Step 2: Based on the xsBert model, feature encoding is performed on the fact description part to obtain semantic encoding information of the fact description, and bidirectional feature extraction is performed on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description;

[0098] Step 3: Construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and use the fully connected layer to map the fact description feature representation to the crime and legal provisions to obtain the probability value of the judicial case to be predicted on each crime and legal provision;

[0099] Step 4: Perform threshold judgment on the probability value to obtain a judgment prediction result.

[0100] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the judgment of a judicial case, characterized in that: include: Obtaining original data of judicial judgment documents, and preprocessing the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part, and a legal article judgment part; Based on the xsBert model, feature encoding is performed on the fact description part to obtain the semantic encoding information of the fact description, and bidirectional feature extraction is performed on the semantic encoding information of the fact description based on the BiGRU model to obtain the feature representation model of the fact description; Construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and use the fully connected layer to map the fact description feature representation to the crime and legal provisions to obtain the probability value of the judicial case to be predicted on each crime and legal provision; A threshold value is judged on the probability value to obtain a judgment prediction result.

2. A method for predicting a judicial case decision according to claim 1, characterized in that: The method of obtaining the original data of the judicial judgment document and preprocessing the original data to obtain the judgment data set includes: Obtaining original data of judicial decision documents; Based on the fact description keywords, the crime judgment keywords and the legal article judgment keywords, the fact description part, the crime judgment part and the legal article judgment part are extracted respectively; Extract crime labels and legal article labels from the crime judgment part and the legal article judgment part to form a label list; Set the specified crime label and the specified legal article label to filter out the original data that does not contain the specified crime label and the specified legal article label; Construct a judgment dataset of <fact description content, crime label set, and legal provision label set>.

3. A method for predicting the judgment of a judicial case according to claim 2, characterized in that: The method of performing feature encoding on the fact description part based on the xsBert model to obtain semantic encoding information of the fact description, and performing bidirectional feature extraction on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description includes: Obtain the fact description content of the judgment data set, insert the [CLS] symbol at the first position of each fact description sentence in the fact description content, and then input it into the xsBert model for encoding; Obtain the semantic feature vector of each fact description sentence based on the xsBert model; The semantic feature vector is mapped to the same latent space based on the BiGRU network, the fine-grained features of the text are extracted, and a fact description feature representation model is obtained according to the fine-grained features of the text.

4. A method for predicting the judgment of a judicial case according to claim 3, characterized in that: The network structure of the xsBert model includes an input layer, a hidden layer and an output layer; The number of hidden layers is 12 in total, and the hidden layers include an attention layer, a fully connected layer, an intermediate layer, and a hidden output layer; The input layer obtains the word vector of the fact description sentence as an input vector; The attention layer is a multi-head attention mechanism with 12 heads; After the attention layer obtains the attention score of the fact description sentence, it superimposes the input vector as output; The output of the attention layer is connected to a fully connected layer, and then the output of the middle layer is obtained through the activation function GELU; The hidden output layer performs full connection and Dropout operations on the output of the intermediate layer to obtain the output of the hidden output layer; The output of all hidden output layers is used as the output of the xsBert model.

5. A method for predicting judgment of a judicial case according to claim 3, characterized in that: The method of mapping the semantic feature vector to the same hidden space based on the BiGRU network, extracting the fine-grained features of the text, and obtaining the fact description feature representation model according to the fine-grained features of the text includes: The semantic feature vectors are fused to obtain a fusion vector; Initialize a bidirectional GRU network and input the fusion vector into the bidirectional GRU network to capture the fine-grained features of the text; Concatenate the output representations of the last two time steps in the bidirectional GRU network to obtain the final fact description feature representation model.

6. A method for predicting the judgment of a judicial case according to claim 5, characterized in that: The text fine-grained feature output is expressed as: output=GRU(fact_all,num_layer,Dim,bidirectional,dropout) Among them, Dim is the model encoding dimension, num_layer is the number of network layers; fact_all represents the fusion vector; bidirectional indicates whether it is bidirectional encoding; dropout represents the probability of random discarding to prevent overfitting.

7. A method for predicting judgment of a judicial case according to claim 2, characterized in that: The case description label data set of the case to be predicted is constructed, the case description label data set is input into the fact description feature representation model, and the fact description feature representation is mapped to the crime name and legal provision by the fully connected layer to obtain the probability value of the judicial case to be predicted on each crime name and legal provision; including method: Obtain the original data of the case to be predicted and construct a case description label dataset of <fact description content, crime label set, and legal article label set>; The case description label data set is input into the fact description feature representation model, and the features are mapped to legal provisions and crimes through a fully connected layer to obtain the distribution probability of the case to be predicted in each legal provision label and crime label; The distribution probability of the case to be predicted in each legal article label and crime label is input into the Sigmoid activation function to calculate the probability value of the judicial case to be predicted in each crime and legal article.

8. A method for predicting the judgment of a judicial case according to claim 7, characterized in that: The method of performing threshold judgment on the probability value to obtain the judgment prediction result includes: mapping the probability values ​​of the judicial case to be predicted on each crime and legal provision into the real crime and legal provision through a hash table.

9. A judgment prediction system for judicial cases, characterized in that: A method for predicting a judgment of a judicial case according to any one of claims 1 to 8, the system comprising: A preprocessing module is used to obtain the original data of the judicial judgment document and preprocess the original data to obtain a judgment data set; the judgment data set includes a fact description part, a crime judgment part and a legal article judgment part; A feature extraction module is used to perform feature encoding on the fact description part based on the xsBert model to obtain semantic encoding information of the fact description, and to perform bidirectional feature extraction on the semantic encoding information of the fact description based on the BiGRU model to obtain a feature representation model of the fact description; The prediction module is used to construct a case description label dataset of the case to be predicted, input the case description label dataset into the fact description feature representation model, and map the fact description feature representation to the crime name and legal provision by the fully connected layer to obtain the probability value of the judicial case to be predicted on each crime name and legal provision; The output module is used to perform threshold judgment on the probability value to obtain a judgment prediction result.

10. A computer readable medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement a method for predicting a decision in a judicial case as described in any one of claims 1 to 8.

Images