Sentencing Methods for Economic Crimes Involving Multiple Individuals Based on Case Fusion and Two-Tier Attention

Abstract

The present application relates to a kind of based on case fusion and double-layer attention of the method for sentencing of economic crime of stakeholder, belong to natural language processing and judicial information technology field.The present application constructs case library by entity identification to decided case judgment book, and extracts the text features of judicial document of undecided case, i.e.case feature and case feature, learns the semantic information of text feature by the powerful word vector representation capability of bert pre-training model, converts text feature into word vector sequence as model input, using double-layer attention mechanism gives weight to text word and sentence respectively, introduce two-layer structure Bi-GRU neural network as encoder, text feature is encoded and sent into Softmax classifier to realize sentencing prediction, while meeting the sentencing prediction of undecided case, guarantee the accuracy of result.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and judicial case handling information technology, specifically involving a sentencing method for economic crimes involving a large number of people based on case fusion and dual-layer attention. Background Technology

[0002] Economic crimes involving large numbers of people are among the most serious economic crimes currently affecting the public, with significant social impact, huge sums of money involved, and a major impact on social stability. They are also characterized by a large number of suspects at various levels. Sentencing suspects in such cases requires consideration of the proportionate nature of the crime to the offense, as well as the judgments of similar cases. Therefore, matching similar cases with the suspect's level is a unique sentencing requirement for these crimes. It is necessary to find cases of similar severity levels in the same region or economically comparable areas, such as those involving the amount of illegal fundraising. Then, the suspect's level is matched based on their criminal acts, circumstances, position, and employment period. Precedents for suspects at the same level are found for reference. Manually finding similar cases at multiple levels and calculating sentences in accordance with regulations is difficult. To assist prosecutorial work, a sentencing method consistent with the characteristics of economic crimes involving large numbers of people is designed to scientifically calculate sentences for suspects with different levels of criminal responsibility. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] The technical problem this invention aims to solve is how to provide a sentencing method for economic crimes involving multiple parties based on case fusion and dual-layer attention, in order to address the problem that existing sentencing methods for economic crimes involving a large number of people, huge sums of money, and complex circumstances awaiting judgment, often have a single dimension of case features that makes it difficult to guarantee accuracy.

[0005] (II) Technical Solution

[0006] To address the aforementioned technical problems, this invention proposes a sentencing method for economic crimes involving multiple parties based on case fusion and dual-layer attention. This method includes the following steps:

[0007] S1. Develop rule templates based on document type, cause of action, sentencing circumstances, and judgment result. The rule templates include text classification keywords and entity recognition rules based on text type, cause of action, common sentencing circumstances, and judgment result.

[0008] S2. Select the corresponding rule template in S1 according to the document type, cause of action and crime, and use the rule template to identify case information, defendant sentencing circumstances and judgment results in the judgments of adjudicated cases in order to build a case database.

[0009] S3. Select the corresponding rule template in S1 according to the document type, cause of action, and crime of the pending case, and complete the identification of case feature entities of pending case information and defendant sentencing circumstances;

[0010] S4. Based on the case characteristics extracted from the documents of the pending cases, query the case database for sentencing judgments of adjudicated cases that match similar case characteristics, and use these as case characteristics for the pending cases.

[0011] S5. Introduce the BERT Chinese pre-trained model, which integrates the pending case features extracted in S3 with the similar case features extracted in S4 into pending case text features and transforms them into word vector sequences, which are used as inputs to the model's encoding layer.

[0012] S6. The word vector sequence input in S5 is fed into the encoding layer. A Bi-GRU (bidirectional gated recurrent unit) neural network model is selected as the encoding layer. The basic information vector of the case, the sentencing circumstances vector, and the similar case judgment result vector are generated by combining word-level and sentence-level attention context vectors.

[0013] S7. Concatenate the three text vectors generated in S6 and input them into the Softmax classifier. The classifier calculates the probability distribution of different sentence intervals to obtain the sentencing prediction results for the pending cases.

[0014] (III) Beneficial Effects

[0015] This invention proposes a sentencing method for economic crimes involving multiple suspects based on case fusion and a two-layer attention recurrent neural network. The method aims to predict sentencing for pending cases by constructing a database of adjudicated cases, applying case fusion technology to cases awaiting adjudication, and using a two-layer attention recurrent neural network to encode the fused text features. This provides sentencing references for cases involving multiple suspects and complex circumstances, reducing the extensive manual case analysis work for judges and prosecutors, assisting in sentencing decisions, and improving the efficiency of case handling.

[0016] This invention proposes a sentencing method for economic crimes involving multiple parties based on case fusion and a two-layer attention recurrent neural network. It constructs a case database from judgments of adjudicated cases through entity recognition, and extracts textual features (case features and similar features) from judicial documents of pending cases. The semantic information of these textual features is learned through the powerful word vector representation capabilities of a BERT pre-trained model, converting the textual features into word vector sequences as model input. A two-layer attention mechanism is used to assign weights to words and sentences in the text. A two-layer Bi-GRU neural network is introduced as an encoder to encode the textual features and feed them into a Softmax classifier to achieve sentencing prediction. This method satisfies the sentencing prediction requirements of pending cases while ensuring the accuracy of the results. Attached Figure Description

[0017] Figure 1 This is a business logic diagram of the sentencing method for economic crimes involving multiple parties based on case fusion and a two-layer attention recurrent neural network, as described in this invention.

[0018] Figure 2 This is a structural diagram of a recurrent neural network model based on case fusion and two-layer attention. Detailed Implementation

[0019] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0020] The technical problem to be solved by this invention is to provide a sentencing method for economic crimes involving a large number of people based on case fusion and a two-layer attention recurrent neural network, so as to solve the problem that existing sentencing methods have a single dimension of case feature extraction and cannot guarantee accuracy in cases involving a large number of people, huge amounts of money, and complex circumstances awaiting judgment in the field of economic crimes.

[0021] This invention relates to a sentencing method for economic crimes involving multiple parties, based on case fusion and a two-layer attention recurrent neural network, belonging to the fields of natural language processing and judicial case information technology. This invention constructs a case database from judgments of adjudicated cases through entity recognition, and extracts textual features from judicial documents of pending cases, namely case features and similar case features. It learns the semantic information of textual features through the powerful word vector representation capabilities of a BERT pre-trained model, converting the textual features into word vector sequences as model input. A two-layer attention mechanism is used to assign weights to words and sentences in the text. A two-layer Bi-GRU neural network is introduced as an encoder to encode the textual features and feed them into a Softmax classifier to achieve sentencing prediction. This method satisfies the sentencing prediction requirements of pending cases while ensuring the accuracy of the results.

[0022] To address the aforementioned technical problems, this invention proposes a sentencing method for economic crimes involving multiple perpetrators based on case fusion and a two-layer attention recurrent neural network. The method includes the following steps:

[0023] S1. Develop rule templates based on document type, cause of action, sentencing circumstances, and judgment result. The rule templates include text classification keywords and entity recognition rules based on text type, cause of action, common sentencing circumstances, and judgment result.

[0024] S2. Select the corresponding rule template in S1 according to the document type, cause of action and crime, and use the rule template to identify case information, defendant sentencing circumstances and judgment results in the judgments of adjudicated cases in order to build a case database.

[0025] S3. Select the corresponding rule template in S1 based on the document type (indictment, review report) and cause of action of the pending case, and complete the identification of case feature entities of pending case information and defendant sentencing circumstances;

[0026] S4. Based on the case characteristics extracted from the documents of the pending cases, query the case database for sentencing judgments of adjudicated cases that match similar case characteristics (case information, sentencing circumstances of the defendant), and use these as the case characteristics of the pending cases.

[0027] S5. Introduce the BERT Chinese pre-trained model, which integrates the pending case features extracted in S3 with the similar case features extracted in S4 into pending case text features and transforms them into word vector sequences, which are used as inputs to the model's encoding layer.

[0028] S6. The word vector sequence input in S5 is fed into the encoding layer. A Bi-GRU (bidirectional gated recurrent unit) neural network model is selected as the encoding layer. The basic information vector of the case, the sentencing circumstances vector, and the similar case judgment result vector are generated by combining word-level and sentence-level attention context vectors.

[0029] S7. Concatenate the three text vectors generated in S6 and input them into the Softmax classifier. The classifier calculates the probability distribution of different sentence intervals to obtain the sentencing prediction results for the pending cases.

[0030] Example 1:

[0031] This invention relates to a sentencing method for economic crimes involving multiple parties based on case fusion and a two-layer attention recurrent neural network, comprising the following steps: 1. Identifying common sentencing factors in judicial cases and developing entity recognition rule templates based on document type and cause of action; 2. Selecting rule templates to perform entity recognition on judgments of adjudicated cases to construct a case database; 3. Using rule templates to perform entity recognition on cases pending adjudication to extract case features; 4. Using the features of the pending case as a basis, querying similar judgments in the case database as case features for this case; 5. Fusing case features and case features and introducing a BERT Chinese pre-trained model to convert the fused text features into word vector sequences, which are then used as input to the encoding layer model; 6. The encoding layer constructs word-level encoders and sentence-level encoders based on a Bi-GRU neural network, while introducing a two-layer attention mechanism to assign attention weights to word-level and sentence-level text representation vectors, outputting encoded text representation vectors; 7. Feeding the text representation vectors into a classifier to calculate the probability value of the corresponding sentence, using the sentence value with the highest probability value as the sentencing prediction result, realizing sentencing recommendations for pending cases, and intelligently assisting case handlers in making decisions.

[0032] This invention provides a sentencing method for economic crimes involving multiple parties based on case fusion and a two-layer attention recurrent neural network. It introduces case features by constructing a case database and uses a Bi-GRU neural network model with a two-layer attention mechanism to predict sentencing. The process includes:

[0033] Step S1: Based on the "Guiding Opinions on Sentencing for Common Crimes" issued by the Supreme People's Court and the Supreme People's Procuratorate, common sentencing factors in judicial cases are identified (juvenile delinquency, elderly delinquency, attempted crimes, accomplices, surrender, confession, plea bargaining, restitution, meritorious service, etc.). A rule template is then developed based on document type and case cause. This template includes text classification keywords and regular expression entity recognition rules based on text type, case cause, and common sentencing factors. For example, the text classification keywords in a judgment for the crime of illegally absorbing public deposits are: '(Ascertained by the trial | Prosecutor's charges | Public prosecutor's charges | This court's opinion)'. Combined with the subsequent regular expression, entity recognition of case information and defendant information can be achieved.

[0034] Step S2: Based on the type and cause of action of the adjudicated case judgment, select the S1 rule template to complete the identification of the adjudicating court, basic information of the defendant, and sentencing factors in the adjudicated case judgment. After data cleaning, the unstructured text data is transformed into structured text data, thereby constructing a case database.

[0035] Step S3: Based on the common sentencing factors identified in S1, and combined with the document type of the pending case (indictment, review report) and the S1 rule template for selecting the cause of action, realize the case characteristics x of the defendant's basic information, the court of judgment, and common sentencing factors in the pending case. base Entity recognition;

[0036] Step S4: Considering the court that made the judgment and common sentencing factors as case-related elements, and referring to the characteristics of pending cases extracted in S3, query the case-related database for cases with characteristics x. base Similar judgments in already adjudicated cases serve as a case-like characteristic for this case. similar , will x base With x similar Feature fusion is performed to obtain textual features x of the pending case;

[0037] Step S5: Model Input Layer. A BERT-based Chinese pre-trained model is introduced, which has powerful word vector representation capabilities. It learns semantic information from the fused text features x through a Transformer encoder and outputs a word embedding sequence x'. x' is used as the input to the model's encoding layer, improving the semantic understanding of the contextual information of legal documents.

[0038] Step S6: Model encoding layer.

[0039] Text Encoder: The S5 word vector sequence x' is a set of sequences with a two-layer structure: one is a set of multiple sentence sequences that make up the case text, and the other is a set of multiple word sequences that make up each sentence. Therefore, a Bi-GRU neural network model with a two-layer encoder structure is constructed as the encoder to learn the vector representation of the case text features. The two-layer encoder structure includes a word-level encoder and a sentence-level encoder. Assume the word vector sequence x' = [x1, x2, ..., x...]. N ], where the nth element is represented as x n (n∈[1,N]), for element x n GRU forward hidden state g fn With GRU backward hidden state g bn They are respectively:

[0040] h fn =GRU([x1,x2,…,x n ])

[0041] h bn =GRU([x N ,x N-1 ,…,x n ])

[0042] Concatenate vectors h n =[hfn ,h bn [This is] the output of the Bi-GRU neural network.

[0043] A two-layer attention mechanism is introduced: a global contextual word-level attention vector w and a sentence-level attention vector s are used to adjust the encoder output vector h. n Calculate word-level and sentence-level representation vectors separately.

[0044] h n The vector contains three sub-vectors: a vector of basic case information, a vector of the defendant's criminal circumstances, and a vector of similar case judgment results. Based on the two-layer attention mechanism, targeting h... n The three seed vectors are respectively introduced into the word-level attention context vector w b w f and w r Generate multiple sentence vector sequences, and then combine them with sentence-level attention context vectors s. b s f and s r Generate three types of text vectors, namely, the basic case information vector d. b Sentencing factors vector d f and the vector d of similar case judgment results r . Take the three vectors d b d f and d r The concatenation results in a case representation vector d:

[0045] d = [d b d f d r ]

[0046] Step S7: Model Output Layer. The case representation vector d from S6 is fed into the Softmax classifier. The Softmax classifier calculates the probability p of belonging to each sentence label. i p i ∈[0,1], i∈[1,I], where p i The probability that the sentence of the pending case falls within this range is represented by p, where I represents the total number of label types for dividing the sentence range. i The maximum possible sentence range will be used as the sentencing recommendation for suspects in this type of economic crime involving a large number of people.

[0047] The model chooses cross-entropy as the loss function, and the formula for the loss function Loss F is as follows:

[0048]

[0049] Where r iThis represents a label flag variable, which is 1 when the sentencing prediction result is the same as the actual case label, and 0 otherwise.

[0050] This invention proposes a sentencing method for economic crimes involving multiple suspects based on case fusion and a two-layer attention recurrent neural network. The method aims to predict sentencing for pending cases by constructing a database of adjudicated cases, employing case fusion technology on cases awaiting adjudication, and using a two-layer attention recurrent neural network to encode the fused text features. This provides sentencing references for pending cases with numerous suspects and complex circumstances, reducing the extensive manual case analysis work for judges and prosecutors, assisting in sentencing decisions, and improving the efficiency of case handling.

[0051] This invention proposes a sentencing method for economic crimes involving multiple parties based on case fusion and a two-layer attention recurrent neural network. It constructs a case database from judgments of adjudicated cases through entity recognition, and extracts textual features (case features and similar features) from judicial documents of pending cases. The semantic information of these textual features is learned through the powerful word vector representation capabilities of a BERT pre-trained model, converting the textual features into word vector sequences as model input. A two-layer attention mechanism is used to assign weights to words and sentences in the text. A two-layer Bi-GRU neural network is introduced as an encoder to encode the textual features and feed them into a Softmax classifier to achieve sentencing prediction. This method satisfies the sentencing prediction requirements of pending cases while ensuring the accuracy of the results.

[0052] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for victim-based economic crime sentencing based on case-based fusion and double-layer attention, characterized in that, The method comprises the following steps: S1, according to the type of document, the case, the sentencing circumstances, the judgment result formulating rule template, the rule template includes the entity recognition rule of text classification keyword, regular expression according to the text type, the case, the common sentencing circumstances, the judgment result; S2, according to the type of document, the case, the corresponding rule template of S1 is selected, and the rule template is used to realize the entity recognition of case information, defendant sentencing circumstances and judgment result of decided case judgment book to construct case database; S3, according to the type of document, the case, the corresponding rule template of S1 is selected, and the entity recognition of case characteristics of undecided case information and defendant sentencing circumstances is completed; S4, the case characteristics of undecided case document are extracted as the standard, the sentencing judgment result of decided case of similar case characteristics is inquired from the case database, which is used as the case characteristics of undecided case; S5, the Chinese pre-training model of bert is introduced, the undecided case characteristics extracted by S3 and the case characteristics extracted by S4 are fused into the text characteristics of undecided case and converted into word vector sequence, which is used as the input of model coding layer; S6, the word vector sequence input by S5 is sent to the coding layer, the Bi-GRU (bidirectional gated recurrent unit) neural network model is selected as the coding layer, the word level and sentence level attention context vector are combined to generate case basic information vector, sentencing circumstances vector and case judgment result vector; S7, the three text vectors generated by S6 are spliced and input into Softmax classifier, the probability distribution of different prison term intervals is calculated to obtain the sentencing prediction result of undecided case; Wherein, The S6 specifically comprises: A Bi-GRU neural network model with a two-layer encoder structure is constructed as an encoder to learn the vector representation of the case text features, and the two-layer encoder structure includes a word-level encoder and a sentence-level encoder. It is assumed that the word vector sequence , where the nth element is represented as x n , n∈[1,N], for the element x n , the GRU forward hidden state g fn and the GRU backward hidden state g bn are respectively: Concatenating vectors as output of the Bi-GRU neural network The global context word-level attention vector w and the sentence-level attention vector s are introduced to the encoder output vector The word-level representation vector and the sentence-level representation vector are calculated respectively; The vector contains three sub-vectors containing information: a vector of basic case information, a vector of the defendant's criminal circumstances, and a vector of similar case judgment results; based on the setup of the two-layer attention mechanism, targeting... The three seed vectors are respectively introduced into the word-level attention context vector w b w f and w r Generate multiple sentence vector sequences, and then combine them with sentence-level attention context vectors s. b s f and s r Generate three types of text vectors, namely, the basic case information vector d. b Sentencing factors vector d f and the vector d of similar case judgment results r ; Transform the three vectors d b d f and d r The concatenation results in a case representation vector d: 。 2. The case-based fusion and double-layer attention-based method for quantifying the sentence of economic crimes of the principal party according to claim 1, wherein, The step S2 further comprises: the recognition result is subjected to data cleaning to convert the unstructured text data into structured text data, so as to construct the case database.

3. The case-based fusion and double-layer attention-based method for quantifying the sentence of economic crimes of the principal party according to claim 1, wherein, The type of document in the step S3 comprises: the opinion of prosecution and the review report.

4. The case-based fusion and double-attention-based method for quantifying the sentence of economic crimes of the principal party according to claim 3, wherein, The S3 specifically includes: according to the common sentencing factors combed in S1, combining with the type of the pending case document and the cause of action, selecting S1 rule template, realizing the case characteristics x of the basic information of the defendant in the pending case, the judgment court and the common sentencing factors base Entity recognition.

5. The case-based fusion and double-attention-based method for quantifying the sentence of economic crimes of the principal party according to claim 4, wherein, The similar case characteristics in the S4 refer to similar case information and defendant sentencing circumstances.

6. The case-based fusion and double-layer attention-based method for quantifying the punishment of economic crimes of the subject type according to claim 4, wherein, The S4 specifically includes: considering the decision court and common sentencing circumstances as the case elements, querying the case characteristics x from the case library with the pending case characteristics extracted by the S3 as the reference base The decision results of the similar decided cases as the case elements x of the present case similar .

7. The case-based fusion and double-layer attention-based method for quantifying the punishment of economic crimes of the subject type according to claim 6, wherein, S5 specifically includes: x base With x similar Feature fusion is performed to obtain text features x of the pending case. A BERT-based Chinese pre-trained model is introduced, which learns semantic information from the fused text features x through a Transformer encoder and outputs a sequence of word embeddings. ,Will As input to the model's encoding layer.

8. The case-based fusion and double-attention-based method for quantifying the punishment of economic crimes of the subject type according to claim 7, wherein, The S7 specifically includes: sending the S6 case representation vector d into a Softmax classifier, and the Softmax classifier is calculated to obtain the probability p belonging to each criminal term label i , i p ∈[0,1], i∈[1,I], wherein p i represents the probability that the sentencing of the pending case belongs to the interval, I represents the total number of label types for the division of the criminal term interval, and the criminal term interval with the maximum probability value p i is selected as the sentencing suggestion for the suspect of the type of economic crime.

9. The case-based fusion and double-attention-based method for quantifying the sentence of economic crimes of the principal party according to claim 8, wherein, The model selects cross entropy as the loss function, and the formula of the loss function Loss F is as follows: where r i denotes the label flag variable, which is 1 if the recidivism prediction result matches the true case label, and 0 otherwise.

Images