Law supervision clue generation method based on general information extraction technology

Abstract

The invention relates to a legal supervision clue generation method based on a general information extraction technology, and belongs to the technical field of natural language processing and judicial case handling information. The method comprises the following steps: firstly, constructing a legal supervision database by using electronic evidences of related cases and decisions of determined cases through an entity recognition technology; secondly, text features and case features are extracted through the general information extraction technology, and a law supervision rule module capable of being dragged and edited is formed through the segmented coding technology; and finally, forming a legal supervision model according to the case characteristics of the related crime names of the financial crimes and the constituent elements, and outputting case handling clues by fusing the supervision rule module. According to the method, a worker can flexibly configure examination conditions, and more accurate law supervision query is realized, so that the implicit crime association relationship is mined, and the case handling quality and efficiency are improved.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and judicial case handling information technology, and specifically relates to a legal supervision clue generation method based on general information extraction technology. Background Art

[0002] This paper utilizes general information extraction technology to extract elements from electronic documents, chat logs, and other electronic evidence files in various cases, employing big data thinking to explore the internal relationships within the cases. The legal supervision model designed by this paper can capture and analyze transaction data, interest rate data, and loan relationship data from judgment documents in private lending civil cases, and compare the litigation status of parties involved in private lending disputes. Summary of the Invention

[0003] (1) Technical issues to be resolved

[0004] The technical problem to be solved by the present invention is how to provide a legal supervision clue generation method based on general information extraction technology to solve the problem of discovering missed clues in the field of economic crimes through existing electronic evidence data.

[0005] (2) Technical solution

[0006] In order to solve the above technical problems, the present invention proposes a method for generating legal supervision clues based on general information extraction technology, which includes the following steps:

[0007] S1. Select a data preprocessing template for S2 based on the type of electronic evidence, wherein a text information extraction template based on OCR technology, a structured document processing template, or an object detection template is selected;

[0008] S2. Preprocess the electronic evidence data, use the random forest algorithm to interpolate missing values, identify and delete duplicate data, use the categorical data coding method to convert the data into digital form, and then analyze and classify the data;

[0009] S3. Pre-train representation models based on different types of legal documents and electronic evidence, learn the semantic information of text features through the uie-m-base basic model, and provide semantic knowledge support for the construction of the legal supervision model;

[0010] S4. Based on the case elements extracted from electronic evidence, legal documents, and other supporting materials, the relevant data is cleansed using a rule-based engine to form a legal supervision database;

[0011] S5. Use the rule engine to implement the preset legal supervision model for legal supervision clue generation;

[0012] S6. Through the business-level abstraction of service resources and service virtualization methods, model users can adjust model element selection, review rules, review result configuration and review scope confirmation;

[0013] S7. Use the legal supervision rules generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4.

[0014] (3) Beneficial effects

[0015] The present invention proposes a method for generating legal supervision clues based on universal information extraction technology. The present invention proposes a method for constructing a legal supervision model based on universal information extraction technology (Universal Information Extraction). A legal supervision database is formed by data cleaning and element extraction of electronic evidence data. Through the business-level abstraction and service virtualization methods of service resources, model users can adjust model element selection, review rule adjustment, review result configuration, and review scope, so that staff can flexibly configure review conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a business logic diagram of the method for constructing a legal supervision model based on general information extraction technology of the present invention;

[0017] Figure 2 Constructing a methodology map for a legal monitoring database based on general information extraction techniques;

[0018] Figure 3 A schematic diagram of extracting cases or evidence clues that meet the rules is loaded from the dictionary file in the database. DETAILED DESCRIPTION

[0019] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings and examples.

[0020] The present invention belongs to the field of natural language processing and judicial case handling information technology, and is a method for generating legal supervision clues in the financial field based on universal information extraction technology.

[0021] In order to solve the above technical problems, the present invention proposes a method for constructing a legal supervision model based on universal information extraction technology, characterized in that the method includes the following steps:

[0022] S1. Select a data preprocessing template for S2 based on the type of electronic evidence, wherein a text information extraction template based on OCR technology, a structured document processing template, or an object detection template is selected;

[0023] S2. Preprocess the electronic evidence data, use the random forest algorithm to interpolate missing values, identify and delete duplicate data, use the categorical data coding method to convert the data into digital form, and then analyze and classify the data;

[0024] S3. Pre-train representation models based on different types of legal documents and electronic evidence, learn the semantic information of text features through the uie-m-base basic model, and provide semantic knowledge support for the construction of the legal supervision model;

[0025] S4. Based on the case elements extracted from electronic evidence, legal documents, and other supporting materials, the relevant data is cleansed using a rule-based engine to form a legal supervision database;

[0026] S5. Use the rule engine to implement the preset legal supervision model for legal supervision clue generation;

[0027] S6. Through the business-level abstraction of service resources and service virtualization methods, model users can adjust model element selection, review rule adjustment, review result configuration, review scope confirmation and other functions;

[0028] S7. Use the legal supervision rules generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4.

[0029] Example 1:

[0030] The technical problem to be solved by the present invention is how to provide a method for extracting electronic evidence elements and constructing a legal supervision model based on general information extraction technology, so as to meet the demand for discovering missed clues in the field of economic crimes through existing electronic evidence data.

[0031] In order to solve the above technical problems, the present invention proposes a method for constructing a legal supervision model based on universal information extraction technology, characterized in that the method includes the following steps:

[0032] S1. Select the data preprocessing template for S2 according to the type of electronic evidence, among which the text information extraction template based on OCR technology or the structured file processing template is selected.

[0033] For paper files or unstructured evidence materials, we use a deep learning OCR (Optical Character Recognition) engine based on Tesseract OCR, extract image features through a convolutional neural network (CNN), and use a recurrent neural network (LSTM) for sequence classification. The LSTM classification method is as follows:

[0034] h t =σ(W h ·[x t ,h t-1 ]+b h )

[0035] Where W h is the convolution kernel weight, b h is the bias, σ is the activation function, x t is the input feature, h t-1 is the hidden state at the previous moment.

[0036] For evidence files with clear formats, such as structured documents (such as JSON, XML, and tables), we extract them using template matching and regular expression techniques. The regular expression matching method is as follows:

[0037]

[0038] where X i is a text segment, δ(X i ) is the matching function.

[0039] Tabular data is processed using the target detection template, table boundaries are detected using YOLOv3, and data is extracted using the spatial structure relationship of table cells.

[0040] S2. Preprocess the electronic evidence data, use the random forest algorithm to interpolate missing values, identify and delete duplicate data, use the categorical data encoding method to convert it into digital form, and then analyze and classify the data.

[0041] Data preprocessing is performed on electronic evidence data, and the random forest algorithm is used to interpolate missing values ​​and identify and delete duplicate data. Multiple decision trees are constructed to fill in unclear or missing content in legal documents. The specific steps are as follows: First, model training is performed, using other complete and feature-annotated legal documents as input and the features of missing values ​​as output. A random forest regression model (for continuous values) or a classification model (for categorical values) is trained. Assume there is a dataset containing feature X and target variable Y, where Y contains missing values. The prediction of random forest can be expressed as:

[0042]

[0043] Where T is the number of decision trees, f t (X) is the prediction of the t-th tree.

[0044] One-Hot Encoding (I.e., encoding a categorical data variable with n different categories, each represented by a unique value) is used to convert the categorical variable into n new binary features, indicating whether the variable belongs to that category. In the binary feature vector of each category, the position belonging to that category is marked as 1, and the rest are marked as 0. For legal texts, One-Hot Encoding treats each category separately, without introducing the order relationship between categories, which can avoid incorrect numerical comparisons.

[0045] Finally, K-Means clustering is used for data analysis and classification. This method can handle complex data sets and can be applied to cluster analysis of legal documents or other case data.

[0046] S3. Pre-train the representation model according to the different types of legal documents and electronic evidence, learn the semantic information of text features through the uie-m-base basic model, and provide semantic knowledge support for the construction of the legal supervision model.

[0047] The representation model is pre-trained according to the different types of legal documents and electronic evidence, and the text or data is converted into feature vectors. The potential semantic information of the data is captured through the uie-m-base basic model. The representation model used in the present invention is BERT (Bidirectional Encoder Representations from Transformers), which is used to generate more accurate and context-aware word embedding representations, thereby improving the performance of NLP (Natural Language Processing) tasks. uie-m-base is a pre-trained language model based on the Transformer architecture, which is used to extract common semantic features from text. The model encodes text semantics through a multi-layer attention mechanism to help extract contextual relationships. The attention mechanism in Transformer is

[0048]

[0049] Among them, Q, K, and V are query, key, and value matrices respectively, and d k The uie-m-base model learns the semantic information of text features, providing semantic knowledge support for S7 to automatically and intelligently mine potential clues or discover problems by analyzing legal documents, electronic evidence, case elements and other data;

[0050] S4. Based on the case elements extracted from electronic evidence, legal documents and other auxiliary materials, the relevant data is cleaned based on the rule engine to form a legal supervision database.

[0051] Based on case elements extracted from electronic evidence, legal documents, and other supporting materials, the relevant data is verified and filtered based on user-defined rules using the Drools rule engine. This process is divided into two steps. The first step is compilation, which involves parsing the legal documents using ANTLR 3 grammar according to preset rules such as the cause of the case, suspect, and the amount involved, and checking the grammatical correctness. An intermediate structure "descr" is then generated to describe the rules. The rule description files used in this invention are dskr files and dictionary files dsl files. The data that passes the rule check is then stored in the database, ultimately forming a legal supervision database.

[0052] S5. Use the rule engine to implement the preset legal supervision model for generating legal supervision clues

[0053] The Jess (Java Expert System Shell) rule engine is used to implement a pre-set legal supervision model for generating legal supervision clues. This pre-set model uses the Jess rule language to define business rules. For example, for loan cases, it defines an interest rate that exceeds a mark, and then extracts cases or evidence clues that meet the rules from the database.

[0054] S6. Through the business-level abstraction of service resources and service virtualization methods, model users can adjust model element selection, review rule adjustment, review result configuration, review scope confirmation and other functions.

[0055] Exposing the Jess rule system in S5 through the REST API enables model users to remotely call and adjust the preset legal supervision model in S5, including functions such as factor selection, review rule adjustment, review result configuration, and review scope confirmation, thereby forming a specific legal supervision model that meets the review needs of judicial personnel for specific cases. The adjusted specific legal supervision model can also be set as a template.

[0056] S7. Use the legal supervision rules generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4.

[0057] The specific legal supervision model generated by S6 is used to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4. Among them, case elements are key information points in the case, including the time, place, suspects, case description, evidence, etc. of the case. Element fusion refers to the unified representation and processing of these case elements from different sources or types, so as to facilitate comparison, matching and analysis between different cases. Secondly, the elements in the case are segmented, and the case is divided into different fragments using time, place, event type, etc. as dividing points, and segmented coding is performed. This coding method helps to independently analyze different parts of complex cases and find related parts. Finally, based on the similarity between case elements or legal rules, the correlation between different cases or evidence is judged to find potential clues.

[0058] Example 2:

[0059] With the diversification of social and economic transactions, private lending has become a widely used form of lending, but illegal lending and professional lending have also emerged. Currently, those who engage in illegal lending activities with serious circumstances are convicted and punished as illegal business operations. In practice, such cases still face the following problems:

[0060] 1. Difficulty in identifying issues in individual cases: When a private lending dispute arises between parties, civil litigation may be initiated in accordance with legal provisions. However, during the case handling process, staff are unable to promptly determine whether the party has engaged in other private lending activities, and are unable to determine whether the party has repeatedly or frequently lent funds.

[0061] 2. Difficulty in initiating subsequent procedures: When lenders establish creditor-debtor relationships through civil litigation, it is difficult to initiate subsequent procedures, and it is also difficult to recover stolen goods and losses.

[0062] To address the aforementioned technical issues, this paper proposes a method for generating legal supervision clues based on general information extraction technology. By capturing and cleaning data from judgments in private lending cases and establishing a relevant database, staff can use model building to compare and track the litigation status of parties involved in private lending disputes, and provide early warnings to parties who may be showing signs of being professional lenders.

[0063] S1. Read paper documents such as civil first-instance judgments, mediation documents, and rulings from private lending cases from the past two years, and perform text information extraction and preprocessing using a deep learning engine based on Tesseract OCR. For evidence files with clear formats, such as structured documents (docx, txt), extraction is performed using techniques based on template matching and regular expressions.

[0064] S2. For judgments, mediation documents, and rulings in private lending cases, we use the random forest algorithm to interpolate missing values, identify and delete duplicate data, and use the categorical data encoding method to convert them into digital form. The process is as follows:

[0065] (1) Divide the private lending dataset D into two subsets: one dataset D2 containing missing values, and the other dataset D1 not containing missing values. The Python sample code is as follows:

[0066]

[0067] (2) Use the random forest algorithm to predict the missing values ​​of the "amount" column. In order to use random forest to interpolate missing values, we need to first perform One-Hot Encoding on all categorical variables in the D1 dataset and then train the random forest model to interpolate missing values.

[0068] (3) First, perform One-Hot Encoding on the case type.

[0069]

[0070]

[0071] (4) Then use the one-hot encoding results in D1 and the known amount column to train a random forest regression model, and then use this model to interpolate the missing values ​​of the amount column in D2.

[0072] (5) After completing the missing value interpolation of D2, it is necessary to merge D1 and D2 and perform One-Hot Encoding on the entire dataset and delete duplicate data.

[0073] (6) After One-Hot encoding, the data will be clustered. The present invention will use K-Means clustering (K-Means clustering is a fast clustering method that classifies samples by calculating the distance between data) and assume that the data is to be divided into M categories. The following formulas (1) and (2) are used to form C1, ..., C M Data clustering. The goal of the K-Means algorithm is to minimize the squared error given that the samples are classified into M categories.

[0074]

[0075] where u j is the data cluster C j ,j=1,2,…,M’s mean vector, its expression is

[0076]

[0077] (7) Select the initial cluster centers of the M data clusters, calculate the distances between the other data objects in the legal document dataset D and the cluster centers of each data cluster, and classify the samples according to the minimum distance principle. Then, update the cluster centers of each class, and re-divide the data in the entire dataset into clusters. Repeat the above steps until the cluster centers of each class no longer change, and the algorithm terminates.

[0078] S3. Pre-train the data set representation model based on the characteristics of the loan-related financial crime case in this case.

[0079] (1) Load the pre-trained BERT model and convert the legal document or electronic evidence text into a feature vector

[0080]

[0081]

[0082] (2) Use the uie-m-base model to extract general semantic features. This model is based on a wider range of pre-trained semantic extraction and is specifically used to capture general semantic information in text

[0083]

[0084]

[0085] The uie-m-base model is used to learn the semantic information of text features, providing semantic knowledge support for the construction of legal supervision models;

[0086] S4. This case is based on the legal documents related to "professional lenders". In this example, Drools will be used to check the rule of "loaning to unspecified persons more than 10 times within two years". The data that passes the rule check will then be stored in the legal supervision database.

[0087] (1) Configure the Drools rule engine. First, write Drools rules in Java and integrate Drools through Spring Boot for rule execution and data processing.

[0088] (2) We will define Drools rules to determine whether the user rule "loans are made to unspecified persons more than 10 times within two years" is met.

[0089]

[0090]

[0091] LoanHistory: This entity class represents loan records, including the number of loans and start and end dates.

[0092] loanCount>10: indicates that the number of loans has exceeded 10.

[0093] ChronoUnit.YEARS.between(startDate,endDate)<=2: Determines whether the loan period is within 2 years.

[0094] (3) If the data passes the Drools rule verification, the data is stored in the database. You can use JPA or JDBC to implement the data storage operation.

[0095] S5. Use the Jess (Java Expert System Shell) rule engine to implement the preset legal supervision model for legal supervision clue generation. The professional lender model uses the Jess rule language to define business rules, define the LoanCase class to represent the relevant information of loan cases, and write the Jess rule file rules.clp to define the rule for legal supervision clue generation, "interest rate is higher than 24%." Then, load the dictionary file from the database to extract cases or evidence clues that meet the rule, such as Figure 3 As shown;

[0096] S6. Expose the Jess rule system in S5 through the REST API, enabling model users to remotely call and adjust S5's pre-set legal supervision model's functions, such as factor selection, review rule adjustment, review result configuration, and review scope confirmation. This allows users to create a specific legal supervision model that meets the judicial staff's review needs for specific cases. The adjusted specific legal supervision model can be set as a template. The following is an API example for adjusting the loan amount limit.

[0097]

[0098]

[0099] After API processing, the amountThreshold parameter in the rule will be dynamically adjusted, affecting the screening results when the preset legal supervision model in S5 is executed.

[0100] S7. Use the specific legal supervision rules generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4. The supervision rule for this case is: the same lender has lent money frequently within 2 years. If there are 10 or more records of lending money to different objects within 2 years before the most recent lending date, the lender's information will be included in the focus database. In this example, the case elements are segmented and coded according to different dimensions (such as borrower, lender, interest rate, number of loans). The coding method is as follows:

[0101] Time segmentation: Divide cases according to time nodes, such as coding by year, month, and day.

[0102] Location segmentation: Cases are coded according to the location where they occurred, such as by province, city, district, and other different geographical granularities.

[0103] Event coding: Encode the case type and number of loans to facilitate subsequent comparison.

[0104] Interest rate coding: Compare the interest rate recognized in this case with the LPR of the same year and classify them according to the ratio.

[0105] By matching the results of case element encoding, potential clues between different cases are discovered. The matching method is based on a similarity algorithm (Jaccard similarity). After matching the relevant case elements, the correlation between these cases is further analyzed to output potential clues.

[0106] The present invention proposes a method for constructing a legal supervision model based on universal information extraction technology (Universal Information Extraction). A legal supervision database is formed by data cleaning and element extraction of electronic evidence data. Through the business-level abstraction of service resources and service virtualization methods, model users can adjust model element selection, review rule adjustment, review result configuration, and review scope, so that staff can flexibly configure review conditions.

[0107] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for generating legal supervision clues based on general information extraction technology, characterized in that: The method comprises the following steps: S1. Select a data preprocessing template for S2 based on the type of electronic evidence, wherein a text information extraction template based on OCR technology, a structured document processing template, or an object detection template is selected; S2. Preprocess the electronic evidence data, use the random forest algorithm to interpolate missing values, identify and delete duplicate data, use the categorical data coding method to convert the data into digital form, and then analyze and classify the data; S3. Pre-train representation models based on different types of legal documents and electronic evidence, learn the semantic information of text features through the uie-m-base basic model, and provide semantic knowledge support for the construction of the legal supervision model; S4. Based on the case elements extracted from electronic evidence, legal documents, and other supporting materials, the relevant data is cleansed using a rule-based engine to form a legal supervision database; S5. Use the rule engine to implement the preset legal supervision model for legal supervision clue generation; S6. Through the business-level abstraction of service resources and service virtualization methods, model users can adjust model element selection, review rules, review result configuration and review scope confirmation; S7. Use the legal supervision rules generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4.

2. The method for generating legal supervision clues based on general information extraction technology according to claim 1, characterized in that: In S1, for paper files or unstructured evidence materials, a deep learning OCR engine based on Tesseract OCR is used to extract image features through a convolutional neural network (CNN), and a recurrent neural network (LSTM) is used for sequence classification. The LSTM classification method is as follows: h t =σ(W h ·[x t ,h t-1 ]+b h ) Where W h is the convolution kernel weight, b h is the bias, σ is the activation function, x t is the input feature, h t-1 is the hidden state at the previous moment.

3. The method for generating legal supervision clues based on general information extraction technology according to claim 1, characterized in that: In S1, structured documents are extracted using a technique based on template matching and regular expressions. The regular expression matching method is: where X i is a text segment, δ(X i ) is the matching function.

4. The method for generating legal supervision clues based on general information extraction technology according to claim 1, characterized in that: In S1, the table data is processed using a target detection template, the table boundary is detected using YOLOv3, and the spatial structure relationship of the table cells is used to extract data.

5. The method for generating legal supervision clues based on general information extraction technology according to any one of claims 1 to 4, characterized in that: The S2 specifically includes: Data preprocessing is performed on electronic evidence data. The random forest algorithm is used to interpolate missing values ​​and identify and delete duplicate data. Multiple decision trees are constructed to fill in unclear or missing content in legal documents. The specific steps are as follows: First, model training is performed. Other complete and feature-annotated legal documents are used as input, and the features of missing values ​​are used as output. A random forest regression model or classification model is trained. Assume that there is a data set containing feature X and target variable Y, where Y contains missing values. The prediction of random forest is expressed as: Where T is the number of decision trees, f t (X) is the prediction of the t-th tree; One-Hot Encoding (IHE) is used to convert categorical data into digital form. IHE converts categorical variables into n new binary features, indicating whether a variable belongs to that category. Within each category's binary feature vector, positions belonging to that category are marked as 1, and all others are 0. For legal texts, IHE treats each category independently, eliminating any ordering between categories and thus preventing erroneous numerical comparisons. Finally, K-Means clustering was used for data analysis and classification.

6. The method for generating legal supervision clues based on general information extraction technology according to claim 5, characterized in that: In the S3, The representation model is BERT, which is used to generate more accurate and context-aware word embedding representations, thereby improving the performance of NLP tasks; uie-m-base is a pre-trained language model based on the Transformer architecture, which is used to extract common semantic features from text; this model encodes text semantics through a multi-layer attention mechanism to help extract contextual relationships. The attention mechanism in Transformer is Among them, Q, K, and V are query, key, and value matrices respectively, and d k It is the key dimension; it learns the semantic information of text features through the uie-m-base basic model, and provides semantic knowledge support for S7 to automatically and intelligently mine potential clues or discover problems by analyzing legal documents, electronic evidence, and case element data.

7. The method for generating legal supervision clues based on general information extraction technology according to claim 6, characterized in that: The S4 includes: based on the case elements extracted from electronic evidence, legal documents, and other auxiliary materials, the relevant data is verified and filtered based on user-defined rules based on the Drools rule engine. This process is divided into two steps. The first step is compilation, which is to parse the legal document through ANTLR3 grammar according to preset rules such as the case cause, suspect, and amount involved, and check the correctness of the grammar; then an intermediate structure "descr" is generated to describe the rules. The rule description file is the dskr file and the dictionary file is the dsl file. The data after the rule check is stored in the database, and finally a legal supervision database is formed.

8. The method for generating legal supervision clues based on general information extraction technology according to claim 7, characterized in that: The S5 includes: using the Jess rule engine to implement a preset legal supervision model for generating legal supervision clues. The preset model uses the Jess rule language to define business rules, and then extracts cases or evidence clues that meet the rules from the database.

9. The method for generating legal supervision clues based on general information extraction technology according to claim 8, characterized in that: The S6 includes: exposing the Jess rule system in S5 through RESTAPI, so that model users can remotely call and adjust the element selection, review rule adjustment, review result configuration, and review scope confirmation of the preset legal supervision model in S5, to form a specific legal supervision model that meets the review needs of judicial staff for specific cases, and supports setting the adjusted specific legal supervision model as a template.

10. The method for generating legal supervision clues based on general information extraction technology according to claim 9, characterized in that: The S7 includes: using the specific legal supervision model generated by S6 to perform clue matching based on case element fusion and segmented coding technology in the legal supervision database constructed by S4; wherein, case elements are key information points in the case, including the time, place, suspects, case description and evidence of the case; element fusion refers to the unified representation and processing of these case elements from different sources or types, so as to enable comparison, matching and analysis between different cases; secondly, segmenting the elements in the case, using time, place, and event type as dividing points, dividing the case into different fragments, and performing segmented coding; this coding method helps to independently analyze different parts of a complex case and find related parts; finally, based on the similarity between case elements or legal rules, judge the correlation between different cases or evidence, so as to find potential clues.

Images