Law decision prediction method based on deep extraction network and text structure analysis

Abstract

The invention belongs to the technical field of legal decision prediction, and particularly relates to a legal decision prediction method based on a deep extraction network and text structure analysis. The adopted model comprises an encoder, an attention layer, a depth extraction network and a fact description separation module; coding the fact description and the crime name definition to obtain a fact description feature and a crime name definition feature; performing context representation and an attention mechanism on the fact description features and the crime name definition features to obtain fact description features integrated with legal knowledge; the fact description feature integrated with legal knowledge passes through a deep extraction network, and abstract representation of fact description is obtained and used for predicting a crime name; encoding the predicted criminal name to obtain predicted criminal name features; the context representation of the predicted criminal name features and the fact description passes through a fact description separation module to obtain a law article basis and a criminal period basis; performing feature extraction on the law article basis through another deep extraction network to obtain semantic features of the law article basis for predicting the law article; and the prediction law article is coded to obtain prediction law article features for predicting criminal periods. The significant features and semantic information in the fact description are extracted through the deep extraction network, the dependency relationship of the law article basis and the criminal period basis in the fact description on the text structure is fully extracted, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of legal judgment prediction, and specifically is a legal judgment prediction method based on deep extraction network and text structure analysis. Technical Background

[0002] Legal judgment prediction aims to predict the judgment results of subtasks such as crime, legal provisions, and penalty period based on the description of case facts. The significant features in the fact description play a decisive role in judgment prediction, such as Figure 1 As shown in the figure, the gray background part is the salient features in the fact description. However, when the fact description passes through the deep network, these salient features will be weakened, and the semantic features between the contexts are also difficult to retain.

[0003] In order to utilize the key information in the factual description, some studies have used fine-grained key event information extracted from the factual description to match the event types defined in the legal provisions. This can accurately retrieve the legal provisions applicable to the case, and can use the judgments in the legal provisions to infer the sentence, but ignores the semantic information between the contexts in the factual description. Other studies have used deep neural networks to obtain key information in the factual description for legal judgment prediction, but the feature information in the factual description is easily weakened when passing through the deep neural network, thus affecting the prediction results.

[0004] In actual judicial scenarios, judges usually refer to various parts of the factual description (legal basis and sentence basis) to decide the verdict. Figure 1 As shown in the figure, the solid line and the dotted line are the legal basis and the sentence basis for the judge to make a judgment, respectively. Some studies use generative models to obtain the basis for judgment from factual descriptions, and then make predictions based on the factual descriptions and the generated basis for judgment. Other studies use the results of intermediate subtasks to separate factual descriptions into different environments, and then use them to make predictions for other subtasks.

[0005] The problem with the above method is that it fails to take into account the inherent structure of factual description. For example, the legal basis usually appears before the sentence basis. This inherent structure not only reflects the rigor and standardization of legal document writing, but also contains important prediction information. Ignoring this structure may lead to information omission and reduced prediction accuracy. Therefore, the present invention proposes a legal judgment prediction method based on deep extraction network and text structure analysis. Summary of the invention

[0006] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to propose a legal judgment prediction method based on deep extraction network and text structure analysis.

[0007] The present invention solves the technical problem by adopting the following technical solution:

[0008] A method for predicting legal judgments based on deep extraction networks and text structure analysis, characterized in that the method comprises the following steps:

[0009] Step 1: Build a legal judgment prediction model, including encoder, attention layer, deep extraction network and fact description separation module;

[0010] The fact description and the crime definition are respectively input into the encoder for encoding to obtain the fact description features and the crime definition features; the fact description features and the crime definition features are respectively subjected to the maximum pooling operation to obtain the contextual representation of the fact description and the crime definition; the contextual representation of the fact description and the crime definition is passed through the attention layer to obtain the fact description features that incorporate legal knowledge;

[0011] The fact description features integrated with legal knowledge are input into a deep extraction network to extract the significant features and semantic information in the fact description to obtain an abstract representation of the fact description; the abstract representation of the fact description is passed through the softmax function to obtain the probability distribution of the predicted crime; the probability distribution of the predicted crime is passed through the argmax function to obtain the predicted crime;

[0012] The predicted crime is encoded by an encoder to obtain the predicted crime features; the predicted crime features and the contextual representation of the factual description are passed through the factual description separation module to predict the legal basis and the sentence basis, and the legal basis and the sentence basis are obtained; the legal basis is subjected to feature extraction through another deep extraction network to obtain the semantic features of the legal basis; the semantic features of the legal basis are passed through the softmax function to obtain the probability distribution of the predicted legal provision, and the probability distribution of the predicted legal provision is passed through the argmax function to obtain the predicted legal provision;

[0013] The predicted law is encoded by an encoder to obtain the predicted law feature; the predicted law feature is concatenated with the sentence based on the semantic feature, and then the sentence is predicted by the softmax function to obtain the predicted sentence;

[0014] Step 2: Train the legal decision prediction model and use the trained legal decision prediction model for legal decision prediction.

[0015] Furthermore, the depth extraction network includes a plurality of serially connected depth extraction modules, and the output features of each depth extraction module are spliced ​​to obtain the output features of the depth extraction network; the input features of the depth extraction module are sequentially subjected to upsampling, Inception module A and Inception module B, and the output features of the Inception module B are residually connected with the features after upsampling, and then subjected to K-top pooling operation and full connection layer to obtain the output features of the depth extraction module;

[0016] The feature extraction process of Inception module A is expressed as:

[0017] R A =Concat(C 1×1 (F),C 1×3 (C 1×1 (F)),C 1×7 (C 1×1 (F)),C avg (C 1×1 (F))) (6)

[0018] In the formula, R A is the output feature of Inception module A, F is the input feature of Inception module A, Concat represents the concatenation operation, C 1×1 , C 1×3 and C 1×7 They are 1×1, 1×3 and 1×7 convolution operations, C avg is the average pooling operation;

[0019] The feature extraction process of Inception module B is expressed as:

[0020] R B =Concat(C 1×1 (R A ),C 1×2 (C 1×1 (R A )),C 1×4 (C 1×1 (R A )),C avg (C 1×1 (R A )))(7)

[0021] In the formula, R B is the output feature of Inception module B, C 1×2 , C 1×4 Represent 1×2 and 1×4 convolution operations respectively.

[0022] Furthermore, the fact description separation module includes a Transformer layer and a CRF layer. The Transformer layer uses the predicted crime features as the key vector and value vector, and the contextual representation of the fact description as the query vector, and generates a contextual representation from the crime to the fact through the attention mechanism; the contextual representation from the crime to the fact is feature mapped through the softmax function to obtain the probability distribution of the legal basis and the basis of the sentence; the probability distribution of the legal basis and the basis of the sentence is input into the CRF layer to generate the optimal label sequence of the legal basis and the basis of the sentence; the legal basis and the basis of the sentence are separated from the optimal label sequence of the legal basis and the basis of the sentence.

[0023] Furthermore, the generation process of the fact description feature incorporating legal knowledge is expressed as follows:

[0024]

[0025] Where H is the fact description feature integrated with legal knowledge, is the contextual representation of the factual description, c is the contextual representation of the crime definition, W is the weight coefficient matrix between the factual description and the crime definition, W u is the learnable parameter matrix.

[0026] Furthermore, during the model training process, the training loss is calculated by the following formula:

[0027]

[0028] In the formula, λ j represents the weight factor of subtask j, j = 1, 2, 3 correspond to the three subtasks of law prediction, crime prediction and sentence prediction respectively, K j represents the number of samples of subtask j, y j,k , They represent the true label of the kth sample in subtask j respectively.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. A deep extraction network (MDEN) is designed to extract significant features and semantic information from factual descriptions and legal basis, which helps to improve the accuracy and reliability of predictions. Considering the structural nature of factual descriptions, that is, the dependency between legal basis and sentence basis in the text structure, this structural nature is analyzed through factual description separation, and conditional random fields (CRF) are used to more accurately capture the logical order and interdependence between elements such as legal basis and sentence basis, thereby improving the accuracy of factual description separation and thus improving the accuracy of legal judgment prediction. The combination of deep extraction networks and factual description separation not only improves the prediction accuracy of legal judgments, but also enriches the understanding of the characteristics of legal documents by deep learning technology.

[0031] 2. In the deep extraction module, upsampling is used to enhance the input features, abstract features are learned through convolutional networks, and noise features are filtered using K-top pooling operations. Through in-depth analysis and extraction of significant features and semantic information, the significant features and semantic information in the factual description and legal basis are retained when passing through the deep extraction network, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is an example of a description of the facts of the case;

[0033] Figure 2 is a structural diagram of the legal decision prediction model of the present invention;

[0034] Figure 3 is a structural diagram of the depth extraction network of the present invention;

[0035] Figure 4 is a structural diagram of a depth extraction module of the present invention;

[0036] Figure 5 It is a schematic diagram of the relative positions of the legal basis and the sentencing basis in the description of facts. DETAILED DESCRIPTION

[0037] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to introduce the technical solutions of the present invention in detail and are not intended to limit the protection scope of the present application.

[0038] The present invention provides a legal judgment prediction method based on deep extraction network and text structure analysis (hereinafter referred to as method, see Figures 1 to 5 ), including the following steps:

[0039] Step 1: Build a legal judgment prediction model; the legal judgment prediction model includes an encoder, an attention layer, a deep extraction network, and a fact description separation module;

[0040] The fact description and the crime definition are input into the encoder respectively. The encoder uses the BERT model to encode the fact description and the crime definition to obtain the fact description features and the crime definition features. The encoding process of the fact description is expressed as:

[0041] [h1,...,h n ]=BERT[x1,...,x n ] (1)

[0042] In the formula, [h1,...,h n ] is a sequence of fact description features, BERT is a BERT model, [x1,...,x n ] is a text sequence of fact descriptions, and n is the length of the fact description;

[0043] The fact description features and the crime definition features are subjected to the maximum pooling operation respectively to obtain the contextual representation of the fact description and the crime definition; among them, the contextual representation of the fact description for:

[0044]

[0045] In the formula, maxpooling is the maximum pooling operation;

[0046] The contextual representation of the fact description and the crime definition is input into the attention layer, and the contextual representation of the crime definition is used to query the contextual representation of the fact description, so as to mine the semantic information most relevant to the crime definition in the fact description, and integrate the legal knowledge into the fact description through the attention mechanism, and obtain the fact description features integrated with legal knowledge. The process is expressed as follows:

[0047]

[0048] Where H is the fact description feature integrated with legal knowledge, c is the contextual representation of the crime definition, W is the weight coefficient matrix between the fact description and the crime definition, and W u is a learnable parameter matrix;

[0049] The fact description features integrated with legal knowledge are input into a deep extraction network (MDEN), and the significant features and semantic information in the fact description are extracted to obtain an abstract representation of the fact description. The abstract representation of the fact description encodes multi-granular semantic information and semantic dependencies, and the abstract representation of the fact description is used for crime prediction. The deep extraction network includes multiple deep extraction modules (DEN), 5 in this embodiment; the output features of the previous deep extraction module are the input features of the next deep extraction module, and the output features of each deep extraction module are spliced ​​to obtain the output features of the deep extraction network.

[0050] In the depth extraction module, the input features are sequentially upsampled, passed through Inception module A and Inception module B. The output features of Inception module B are residually connected with the upsampled features, and then passed through K-top pooling and a fully connected layer to obtain the output features of the depth extraction module.

[0051] Upsampling is an enhancement of input features, which facilitates the subsequent extraction of richer feature information. Upsampling is only a copy of the input features, so it does not increase the number of model parameters. The upsampling operation is expressed by the following formula:

[0052] [H i1 ,...,H iv ]=Upsample(H i )(5)

[0053] In the formula, H i is the input feature of the i-th deep feature extraction module, [H i1 ,-,H iv ] is the up-sampled feature, H i1 =H iv =H i , v represents the predefined parameters of upsampling, and Upsample represents the upsampling operation.

[0054] Inception module A and Inception module B are designed to extract features of different granularities from input data using convolutions of different scales. The input features of Inception module A are extracted through four parallel 1×1 convolutions, and the output features of the three 1×1 convolutions are respectively average pooled, 1×3 convolutions, and 1×7 convolutions, and then concatenated with the output features of the last 1×1 convolution to provide a richer feature representation for the next module, and obtain the output features of Inception module A; the feature extraction process of Inception module A is expressed as:

[0055] R A =Concat(C 1×1 (F),C 1×3 (C 1×1 (F)),C 1×7 (C 1×1 (F)),C avg (C 1×1 (F)))(6)

[0056] In the formula, R A is the output feature of Inception module A, Concat represents the concatenation operation, C 1×1 , C 1×3 and C 1×7They are 1×1, 1×3 and 1×7 convolution operations respectively; C avg is an average pooling operation, which is used to reduce the feature dimension and retain important information; F = [H i1 ,...,H iv ] is the input feature of the Inception module A, that is, the feature obtained by upsampling.

[0057] Similarly, the input features of Inception module B are extracted through four parallel 1×1 convolutions, where the output features of three 1×1 convolutions are respectively subjected to average pooling, 1×2 convolution, and 1×4 convolution, and then concatenated with the output features of the last 1×1 convolution to obtain the output features of Inception module B. The feature extraction process of Inception module B is expressed as:

[0058] R B =Concat(C 1×1 (R A ),C 1×2 (C 1×1 (R A )),C 1×4 (C 1×1 (R A )),C avg (C 1×1 (R A )))(7)

[0059] In the formula, R B is the output feature of Inception module B, C 1×2 , C 1×4 Represent 1×2 and 1×4 convolution operations respectively;

[0060] In order to further enhance the feature transfer capability of the model and prevent the gradient vanishing problem during training, the output features of the Inception module B are connected to the features obtained by upsampling through residual connections, which enables rapid information flow between features at different levels, thereby helping to improve the training effect and accelerate the convergence process; the residual connection is expressed as:

[0061] R=F+R B (8)

[0062] Where R is the feature obtained by residual connection.

[0063] The features obtained by residual connection are subjected to K-top pooling operation and full connection layer to obtain the output features of the deep extraction module. The process is expressed as:

[0064] H i '=σ(K-top(R))(9)

[0065] In the formula, H i ' is the output feature of the i-th deep extraction module, σ represents the fully connected layer, and K-top represents the K-top pooling operation.

[0066] The abstract representation of the fact description is passed through the softmax function to obtain the probability distribution of the predicted crime; the probability distribution of the predicted crime is passed through the argmax function to obtain the predicted crime; the process is expressed as follows:

[0067]

[0068] In the formula, is the probability distribution of predicting the crime, softmax is the softmax function, Q is the abstract representation of the fact description, and W c , b c is the learnable parameter matrix and bias, argmax is the argmax function, It is to predict the crime;

[0069] Using encoders to predict crimes Encode and obtain the predicted crime characteristics Considering the text structure of the fact description, the crime characteristics will be predicted and the contextual representation of the fact description Input it into the fact description separation module to predict the legal basis and sentence basis, and obtain the legal basis H a and the sentence is based on H t , used for the prediction of laws and crimes.

[0070] The fact description separation module includes the Transformer layer and the CRF layer. The Transformer layer uses the predicted crime features as the key vector and value vector, and the contextual representation of the fact description as the query vector. It uses the inherent association between the crime definition and the fact description to generate the contextual representation from the crime to the fact. Its operation is expressed as follows:

[0071]

[0072] Where D is the context representation from the crime to the fact, Attention represents the attention mechanism, K and V are the key vector and value vector, Q is the query vector, and d k are the dimensions of the key and value vectors;

[0073] Use the softmax function to map the contextual representation from the crime to the fact into the feature space, and predict the probability distribution of the legal basis and the sentence basis;

[0074] P = softmax(W s D+b s )(13)

[0075] Where P is the probability distribution of legal basis and sentence basis, W s , b s is the learnable parameter matrix and bias;

[0076] The probability distribution P of the legal basis and the sentence basis is input into the CRF layer to generate the optimal label sequence of the legal basis and the sentence basis; the CRF layer can learn the transition relationship between labels, and capture the dependency relationship in the label sequence by considering the transition probability between labels and the conditional probability of each label. For example, the legal basis should be after the sentence basis; the generation process of the optimal label sequence is expressed as:

[0077]

[0078] In the formula, The optimal label sequence representing the legal basis and the sentence basis; is an element of the state transition matrix, representing the transition from the previous label y i-1 Go to current label y i The state transition matrix is ​​the learnable parameter matrix of the CRF layer, which captures the transition probability between adjacent labels in the label sequence; Indicates the label y at position i i The conditional probability of

[0079] The optimal label sequence based on legal provisions and sentence terms Separate from the legal basis H a and the sentence is based on H t ; Legal basis H a After feature extraction by another deep extraction network (MDEN), the legal clause semantic feature H encoding multi-granularity semantic features and semantic dependency relationships is obtained. a '; The legal provisions are based on semantic features H a 'The probability distribution of the predicted law is obtained through the softmax function, and the probability distribution of the predicted law is obtained through the argmax function. The process is expressed as:

[0080]

[0081] In the formula, is the prediction law, W a , b a is the learnable parameter matrix and bias, It is a predictive law;

[0082] Use the encoder to encode the predicted legal provisions to obtain the predicted legal provisions features The predicted legal feature and the sentence semantic feature H t 'Concatenate, and then use the softmax function to predict the sentence to get the predicted sentence; the process is expressed as:

[0083]

[0084] In the formula, is the predicted sentence, W t , b t is a learnable parameter matrix and bias, [;] represents a concatenation operation;

[0085] Step 2: Train the legal judgment prediction model, calculate the training loss according to the following formula, and use the trained legal judgment prediction model for legal judgment prediction;

[0086]

[0087] In the formula, λ j represents the weight factor of subtask j, j = 1, 2, 3 correspond to the three subtasks of law prediction, crime prediction and sentence prediction respectively, K j represents the number of samples of subtask j, y j,k , They represent the true label of the kth sample in subtask j respectively.

[0088] Example

[0089] This example is tested on the CAIL-big and CAIL-small datasets. First, the dataset is preprocessed to filter out infrequent crimes and legal provisions, retain only legal provisions with a frequency greater than 100, and divide the sentences into non-overlapping intervals to filter out multi-label samples.

[0090] Through in-depth analysis of the CAIL-small dataset, the structure of the description of legal facts is shown. Since there is no corresponding labeled data, this paper randomly extracts 50 cases from CAIL-small and marks the legal basis and the sentence basis respectively. Figure 5 The figure shows the relative positions of the two bases in the fact description. The horizontal axis of 0% indicates the beginning of the fact description, and 100% indicates the end of the fact description. The vertical axis indicates the number of fact descriptions. In the figure, blue and yellow respectively represent the legal basis and the sentencing basis in the fact description. For example, the blue area is located between 0% and 70% of a case, indicating that the legal basis is located between 0% and 70% of the fact description. As can be seen from the figure, the relative positions of the legal basis and the sentencing basis in the fact description are fixed. This result shows that the structural nature of the fact description can be considered to separate the fact description.

[0091] For the encoder (BERT model), the maximum length of the fact description is set to 512. To optimize the parameters, the Adam optimizer is used, the learning rate is set to 0.001, the dropout rate is set to 0.5, and the legal judgment prediction model is implemented using PyTorch. 2×H800 GPUs are used for training on the server, and each model is trained for 16 epochs with a batch size of 128.

[0092] In order to verify the effectiveness of the legal judgment prediction model of the present invention, Word2Vec+SVM is selected [1-2] 、FAL [3] ,TOPJUDGE [4] 、LADAN [5] 、NeurJudge [6] , CTM [7] and EMP [8] It is used as a baseline model for comparison, and four indicators including accuracy (Acc), macro precision (MP), macro recall (MR) and macro F1 score are selected for evaluation. Tables 1 and 2 show the experimental results on the CAIL-small dataset and CAIL-big dataset.

[0093] Table 1 Comparison of experimental results between the proposed model and the baseline model on the CAIL-small dataset

[0094]

[0095]

[0096] Table 2 Comparison of experimental results between the proposed model and the baseline model on the CAIL-big dataset

[0097]

[0098] From the above results, it can be seen that all models perform better on the CAIL-big dataset than the CAIL-small dataset. This is because the scale of the CAIL-big dataset is larger than that of the CAIL-small dataset. Larger datasets usually contain more diverse samples, which helps the model learn more extensive and complex feature representations, so that it can better generalize to unseen data. Compared with the baseline model, the indicators of the present invention on the three tasks have been improved to a certain extent, maintaining relatively stable precision and recall rates, especially the F1 score is the highest on the three tasks, indicating that the model has good robustness and generalization ability. Although the accuracy of all models on the sentence prediction task is not very high, the model of the present invention is still the best (Acc is 40.27% and 61.53% respectively), which shows that sentence prediction may be a more complex and challenging task. The present invention analyzes the structurality of fact description through fact description separation, can identify and understand the logical relationship and organizational structure of each part of the fact description, can effectively capture the legal basis and sentence basis, improve the accuracy of judgment prediction, and at the same time improve the interpretability of judgment prediction. The deep extraction network can extract key information and semantic dependencies from the text, further improving the prediction accuracy.

[0099] In order to verify the effectiveness of the legal knowledge fusion, deep extraction network and fact description separation of the present invention, ablation experiments were performed on these three parts on the CAIL-small dataset, and the results are shown in Table 3.

[0100] Table 3 Ablation experiment results on CAIL-small dataset

[0101]

[0102]

[0103] In Table 3, -Att means that the fact description is not integrated with the crime definition, and the fact description is directly used for prediction; -MDEN means that the deep extraction network is not used to extract significant features and semantic information; -Separation means that the fact description is not separated. The ablation experiment results show that after removing Att or MDEN, each indicator in the three tasks has decreased to varying degrees, verifying the effectiveness of the fusion of fact description and crime definition, significant features and semantic information extraction. After removing Separation, each indicator in the two subtasks of legal provision prediction and sentence prediction has decreased to varying degrees, verifying the effectiveness of fact description separation for legal provision and sentence prediction.

[0104] The sources of the above models are as follows:

[0105] [1]Mikolov T,Sutskever I,Chen K,et al.Distributed representations ofwords and phrases and their compositionality[J].Advances in neuralinformation processing systems,2013,26.

[0106] [2]Suykens J A K,Vandewalle J.Least squares support vector machineclassifiers[J].Neural processing letters,1999,9:293-300.

[0107] [3]Luo B,Feng Y,Xu J,et al.Learning to predict charges for criminalcases with legal basis[J].

[0108] arxiv preprint arxiv:1707.09168,2017.

[0109] [4]Zhong H,Guo Z,Tu C,et al.Legal judgment prediction via topologicallearning[C] / / Proceedings of the 2018conference on empirical methods innatural language processing.2018:3540-3549.[5]Xu N,Wang P,Chen L,etal.Distinguish confusing law articles for legal judgment prediction[J].

[0110] arxiv preprint arxiv:2004.02557,2020.

[0111] [6]Yue L,Liu Q,Jin B,et al.Neurjudge:A circumstance-aware neural framework for legal judgment prediction[C] / / Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval.2021:973-982.

[0112] [7]CHEN H,CAID,DAI W,et al.Charge-Based Prison Term Prediction with Deep Gating Network[C / OL] / / Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing(EMNLP-IJCNLP),2019.

[0113] [8]Feng Y,Li C,Ng V.Legal judgment prediction via event extraction with constraints[C] / / Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).2022:648-664.

[0114] Matters not described in this invention are applicable to the prior art.

Claims

1. A legal judgment prediction method based on deep extraction network and text structure analysis, characterized in that: The method comprises the following steps: Step 1: Build a legal judgment prediction model, including encoder, attention layer, deep extraction network and fact description separation module; The fact description and the crime definition are respectively input into the encoder for encoding to obtain the fact description features and the crime definition features; the fact description features and the crime definition features are respectively subjected to the maximum pooling operation to obtain the contextual representation of the fact description and the crime definition; the contextual representation of the fact description and the crime definition is passed through the attention layer to obtain the fact description features that incorporate legal knowledge; The fact description features integrated with legal knowledge are input into a deep extraction network to extract the significant features and semantic information in the fact description to obtain an abstract representation of the fact description; the abstract representation of the fact description is passed through the softmax function to obtain the probability distribution of the predicted crime; the probability distribution of the predicted crime is passed through the argmax function to obtain the predicted crime; The predicted crime is encoded by an encoder to obtain the predicted crime features; the predicted crime features and the contextual representation of the factual description are passed through the factual description separation module to predict the legal basis and the sentence basis, and the legal basis and the sentence basis are obtained; the legal basis is subjected to feature extraction through another deep extraction network to obtain the semantic features of the legal basis; the semantic features of the legal basis are passed through the softmax function to obtain the probability distribution of the predicted legal provision, and the probability distribution of the predicted legal provision is passed through the argmax function to obtain the predicted legal provision; The predicted law is encoded by an encoder to obtain the predicted law feature; the predicted law feature is concatenated with the sentence based on the semantic feature, and then the sentence is predicted by the softmax function to obtain the predicted sentence; Step 2: Train the legal decision prediction model and use the trained legal decision prediction model for legal decision prediction.

2. The legal judgment prediction method based on deep extraction network and text structure analysis according to claim 1 is characterized in that: The deep extraction network includes a plurality of serially connected deep extraction modules, and the output features of each deep extraction module are spliced ​​to obtain the output features of the deep extraction network; the input features of the deep extraction module are sequentially subjected to upsampling, Inception module A and Inception module B, and the output features of the Inception module B are residually connected with the features after upsampling, and then subjected to K-top pooling operation and full connection layer to obtain the output features of the deep extraction module; The feature extraction process of Inception module A is expressed as: R A =Concat(C 1×1 (F),C 1×3 (C 1×1 (F)),C 1×7 (C 1×1 (F)),C avg (C 1×1 (F))) (6) In the formula, R A is the output feature of Inception module A, F is the input feature of Inception module A, Concat represents the concatenation operation, C 1×1 , C 1×3 and C 1×7 They are 1×1, 1×3 and 1×7 convolution operations, C avg is the average pooling operation; The feature extraction process of Inception module B is expressed as: R B =Concat(C 1×1 (R A ),C 1×2 (C 1×1 (R A )),C 1×4 (C 1×1 (R A )),C avg (C 1×1 (R A ))) (7) In the formula, R B is the output feature of Inception module B, C 1×2 , C 1×4 Represent 1×2 and 1×4 convolution operations respectively.

3. The legal judgment prediction method based on deep extraction network and text structure analysis according to claim 1 or 2, characterized in that: The fact description separation module includes a Transformer layer and a CRF layer. The Transformer layer uses the predicted crime features as the key vector and value vector, and the contextual representation of the fact description as the query vector, and generates a contextual representation from the crime to the fact through the attention mechanism; the contextual representation from the crime to the fact is feature mapped through the softmax function to obtain the probability distribution of the legal basis and the basis of the sentence; the probability distribution of the legal basis and the basis of the sentence is input into the CRF layer to generate the optimal label sequence of the legal basis and the basis of the sentence; the legal basis and the basis of the sentence are separated from the optimal label sequence of the legal basis and the basis of the sentence.

4. The legal judgment prediction method based on deep extraction network and text structure analysis according to claim 3 is characterized in that: The generation process of the fact description features incorporating legal knowledge is expressed as follows: Where H is the fact description feature integrated with legal knowledge, is the contextual representation of the factual description, c is the contextual representation of the crime definition, W is the weight coefficient matrix between the factual description and the crime definition, W u is the learnable parameter matrix.

5. The legal judgment prediction method based on deep extraction network and text structure analysis according to claim 1 is characterized in that: During model training, the training loss is calculated by the following formula: In the formula, λ j represents the weight factor of subtask j, j = 1, 2, 3 correspond to the three subtasks of law prediction, crime prediction and sentence prediction respectively, K j represents the number of samples of subtask j, y j,k , They represent the true label of the kth sample in subtask j respectively.

Images