Method and system for extracting knowledge of adjudication rules based on case text

Abstract

The present invention is a method for constructing a rule map with case texts as the knowledge source. It mainly targets the judgment documents of criminal, civil and administrative first-instance, second-instance and retrial cases heard by the court. According to the business content characteristics, judgment document production specifications, semi-structured characteristics, composition characteristics and type characteristics of the judgment documents, an information model of the judgment rule map is designed, and a method for constructing legal terms, a classification method for identifying rules and a method for identifying rule types are proposed. Finally, the construction of the judgment rule map is realized, and an XML structured document is output, so that the computer system can effectively use the judgment rules existing in the case text to carry out legislation, law enforcement and judicial applications, and play a role in the fields of judgment document error correction, judgment document quality review, case quality supervision, and judicial case handling law and regulations.

Description

Technical Field

[0001] The present invention relates to the field of legal knowledge engineering technology, and in particular to a method and system for extracting adjudication rule knowledge based on case texts. Background Art

[0002] Legal rules are behavioral norms that take a certain structural form to specifically stipulate people's legal rights, legal obligations and corresponding legal consequences. Legal rules are generally expressed in natural language in the text of laws and regulations. However, based on the requirements of legal legislation technology, the expression of legal rules in the text of laws and regulations is general and abstract, which is equivalent to the existence of the ontology. In the application of law, it is necessary to operationalize and reason the legal terms in the legal rules for specific scenarios. Especially in the judicial adjudication process, in order to clarify the legal reasoning process of the current case, it is necessary to further operationalize the legal rules into specific rules, which is equivalent to the entity existence of the ontology to match the legal relationship of the current case. Recording this process in the adjudication document will form a more operational adjudication rule. It can be said that the adjudication rule is a more specific and operational entity legal rule. In the adjudication document, the adjudication rule is also expressed in natural language. This knowledge representation method is not well readable for computer systems. It is necessary to re-express the adjudication rules in the case text in the form of a rule map. Therefore, the technology of extracting the adjudication rule map from the case text is the key technology for realizing the computer system to obtain new legal rules. The closest similar technologies and their shortcomings are described as follows:

[0003] 1. A method, device, system, medium and process for constructing a legal knowledge graph based on machine learning (application number / patent number: CN201710339258.8): The core method includes (1) using text features extracted by a natural language understanding module; (2) legal knowledge features learned by a feature machine learning module; (3) legal concepts identified by a knowledge engineering module; (4) identifying legal knowledge points from original legal data; and (5) automatically associating legal knowledge points through a legal concept framework to construct a legal knowledge graph.

[0004] 2. A method for constructing a legal knowledge graph based on information extraction (application number / patent number: CN201910980158.2): The core method includes (1) crawling a large number of legal documents and legal entities from the Internet; (2) extracting legal entities related to law from Baidu Encyclopedia; (3) integrating the crawled legal entities with the legal entities extracted from Baidu Encyclopedia, and linking the legal entities to the legal documents to build a network between the legal entities and the legal documents; (4) constructing an entity network using entity descriptions and entity names, and constructing an entity similarity network based on entity structure information; (5) extracting triples from legal documents and combining them with the legal knowledge network to construct a legal knowledge graph.

[0005] 3. A method for automatically constructing a legal knowledge graph (application number / patent number: CN201710270508.7): (1) Obtain civil, criminal, and administrative trial documents as corpus for constructing a legal knowledge graph, remove stop words from the corpus, and perform word segmentation; (2) Based on the word segmented corpus, extract the subject words of the three types of trial documents, perform part-of-speech tagging and filtering on the extracted subject words, and extract nouns or noun phrase subject words as entity concepts of the legal knowledge graph based on the filtering results; (3) Obtain the subject words associated with each noun or noun phrase extracted. For words that are similar to noun phrases and subject terms, the obtained similar words are tagged and filtered for part of speech. Based on the filtering results, nouns or noun phrases are extracted as entity concepts of the legal knowledge graph; (4) Entity concepts belonging to the same subject terms form a subject relationship. Each subject term entity concept forms a similar relationship with its similar word entity concept. Based on the extracted subject term entity concepts and similar word entity concepts and the relationship between them, the subject term-subject relationship-subject term, subject term-similar relationship-similar word triple structure is formed to construct the legal knowledge graph.

[0006] 4. A method for constructing a judicial judgment rule database, a database and a query method (application number / patent number: CN201910988426.5): The present invention discloses a method for constructing a judicial judgment rule database, a database and a query method, and the construction method includes the following steps: 1) Data collection: by processing case judgment documents, the judgment rules, judgment issues and case elements of the case are obtained, and the cases include cases published online or other cases not published online; 2) Data storage: the judgment issues, judgment rules and case elements are stored in the database and the corresponding relationship between the judgment issues and the judgment rules and case elements is established in the database. The present invention solves the problem that the data of the existing legal database is incomplete and the judgment rules cannot be directly obtained.

[0007] The above four patents have the following shortcomings with respect to the construction of legal rule graphs: (1) There is a lack of targeted solutions for natural language scenarios such as legal rules contained in judicial documents; (2) The generalized legal knowledge graph emphasizes the association system of knowledge and lacks the knowledge representation form of the "rules" that are the core of legal knowledge; (3) They are all general methods that are applicable to the construction of simple legal knowledge graphs, forming graphs based on short-distance semantic associations in text contexts, but no further method design is carried out for the construction of complex knowledge graphs; (4) The graph construction for local text semantics cannot support reasoning between multiple graphs to construct new graphs, because it is difficult to realize the relationship between multiple legal rules to construct secondary graphs; (5) There is no control mechanism for the extraction rate and accuracy rate for each specific type of knowledge graph; (6) The feature machine learning module uses algorithms such as random forests and support vector machines to train and learn legal knowledge features, but does not make good use of the industry characteristics of legal knowledge features, and its training effect is difficult to achieve high-precision indicators. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a method and system for extracting knowledge of judgment rules based on case texts, which can construct a knowledge graph for legal rules.

[0009] In order to solve the above technical problems, the present invention provides a method for extracting knowledge of judicial rules based on case texts, the method comprising: according to the writing specifications and structure of judicial documents and the business standards of litigation procedures, the documents are divided into multi-level text slices from coarse to fine, and a slicing model of the documents is designed; according to the possible slicing positions of judicial rules, based on the results of Chinese text layered slicing, an algorithm model based on rule and deep learning hybrid is adopted to realize the design and algorithm implementation of the slicing model of the natural language segment where the judicial rules are located; according to the business characteristics of the judicial rules, a judicial rule structure model is designed; based on the own legal terminology library, an expert vocabulary is used for the legal text corpus and the judicial document text corpus. The legal terminology database is constructed by using a combination of rule-based bootstrapping and hybrid models; a legal terminology concept map is constructed using the constructed legal terminology database and a deep learning-based relationship extraction model, and the map is stored in standard triples; for the adjudication rule slice corpus identified as "adjudication rules", the legal terminology database is integrated according to different types of features in the adjudication rule model, and a hybrid method based on rules and deep learning is adopted to identify the relationship between legal terms in the "adjudication rule slices", and the adjudication rules existing in the adjudication rule slice corpus are generated according to the adjudication rule structure model, and the link relationship between the adjudication rules and the corpus is recorded; according to the type of extracted legal rules, they are stored in the corresponding legal rule structure model.

[0010] In some embodiments, based on the writing standards and structure of judicial documents and the business standards of litigation procedures, the documents are divided into multiple levels of text pieces from coarse to fine, and a document slicing model is designed, including: summarizing the writing standards and structure of judicial documents, dividing each paragraph of the document into multiple levels of text pieces according to logical relationships, analyzing the first-level text pieces of the judicial documents, including "text header", "basic information segment of the parties", "origin of the case segment", "pleading information segment", "case facts segment", "judgment analysis process segment", "judgment result segment" and "text tail"; designing a document slicing model to store each logical segment of the document, each logical segment contains a number of fine slices. According to the fine slices contained in each paragraph, a document slicing model is designed, such as Figure 2 As shown, each fine slice content is stored in a string type, and the text slice is named according to the content it contains, such as "case fact segment" and "evidence fine slice segment". The entire slicing model is stored in a tree structure.

[0011] In some embodiments, based on the results of Chinese text layered slicing and according to the possible slice positions of the judgment rules, a hybrid algorithm model based on rules and deep learning is used to realize the design and algorithm implementation of the slice model of the natural language segment where the judgment rules are located, including: summarizing the target slice positions where the judgment rules may appear, pointing to the "case fact reasoning section", "evidence authentication section", "fact dispute focus identification section", "judgment essence section", and "legal dispute focus identification section"; based on the business connotation, location characteristics, and semantic characteristics of the judgment rules, a rule-based NLP algorithm model is used to realize the extraction of judgment rule slices in the target slice paragraphs of the case text; the extracted judgment rule slice text is stored in the case xml overall model as a "judgment rule slice" node, and the original text position of the slice is recorded at the same time.

[0012] In some implementations, a structural model of the arbitration rules is designed based on the business characteristics of the arbitration rules, including: designing the structural model of the arbitration rules in combination with the constituent elements of the arbitration rules, the mandatory attributes of the arbitration rules, and the normative content.

[0013] In some embodiments, the structural model of the judgment rules is designed in combination with the constituent elements of the judgment rules, the mandatory attributes of the judgment rules, and the normative content, including: designing a structural model of standard judgment rules, including three groups of elements: behavioral conditions, behavioral patterns, and legal consequences, each group of elements is divided into two types: single factor and multi-factor, supporting two types of storage in the structural model, and designing the logical types between factors in the case of multiple factors, the elements of behavioral conditions, behavioral patterns, and legal consequences include two types: qualitative elements and quantitative elements, the quantitative elements support two modes: quantitative value and quantitative value interval, and one of the three groups of elements in the structural model of the judgment rules, the behavioral conditions and the behavioral patterns, can be omitted; designing a structural model of non-standard judgment rules, a certain behavioral condition or behavioral pattern can also be used as the legal consequence element in the above structural model, and other behavioral conditions and behavioral patterns as two groups of elements in the legal rule structural model constitute a complete judgment rule with this certain behavioral condition or behavioral pattern; designing the mandatory attributes of the judgment rules, including three types: imperative, prohibitive, and guiding; designing the normative content attributes of the judgment rules, including three types: authorization, obligation, and a combination of rights and obligations.

[0014] In some embodiments, based on the own legal terminology database, the legal terminology database is constructed for the legal text corpus and the judicial document text corpus by using the expert vocabulary, the rule-based bootstrap method and the mixed mode, including: preparing the legal text corpus and the case text corpus; generating candidate legal terms; calculating statistical features, respectively calculating the TF-IDF, NC-Value, PMI, left (right) neighbor word entropy and other values ​​of the corpus as statistical features; learning a classifier based on positive and negative samples to learn a binary classifier, and the classifier model can use a decision tree, a random forest or a support vector machine; quality scoring, executing the statistical feature calculation results; using a convolutional neural network, using the previously formed standard library, to train and form new legal term extraction results.

[0015] In some embodiments, preparing legal text corpus and case text corpus includes: dividing the legal text from coarse to fine into multi-level text segments according to the writing standards and text structure of laws and regulations and the legislative technical specifications, designing a slicing model for the legal text, and based on this slicing model, combining the legislative technical features and semantic features of the legal text, and using a rule-based NLP algorithm to slice and store the legal text; designing a legal attribute model based on the legal principles, legislative technical specifications and legal rule application objectives of the laws and regulations, taking the entire legal text as the object, and using the results of Chinese text layered slicing to adopt a hybrid algorithm model based on rules and deep learning to extract and store legal attributes; using a deep learning algorithm model based on the legislative technical specifications and legal article attribute features of the laws and regulations to automatically classify the legal articles, the classification includes at least "legal concept articles", "legal principle articles", "legal rule articles" and "legislative purpose articles", and storing the article classification in the "technical classification" attribute of the legal article.

[0016] In some embodiments, a legal terminology library is constructed and a deep learning-based relationship extraction model is used to construct a legal terminology concept map, which is stored in standard triples, including: designing a relationship type system in the legal terminology concept map, including synonym relationships, abbreviation relationships, hypernym relationships, hyponym relationships, reference relationships, and antonym relationships; extracting based on pattern matching and bootstrapping; expert engineering review to form standard data; deep learning methods and algorithm training; forming a legal terminology concept map based on the legal terminology library.

[0017] In some embodiments, for the judgment rule slice corpus identified as "judgment rules", according to different types of features in the judgment rule model, the legal terminology library is integrated, and a hybrid method of rule-based and deep learning is adopted to identify the relationship between the legal terms in the "judgment rule slice", and according to the judgment rule structure model, the judgment rules existing in the judgment rule slice corpus are generated, and the link relationship between the judgment rules and the corpus is recorded, including: for a single "judgment rule slice" case corpus containing complete judgment rule structure elements, the legal terms are marked, and in combination with the currently marked legal terms and their context, the relationship between the legal terms and the attribution of the legal terms in the judgment rule structure model are identified, and based on this, the judgment rule structure type recognition is implemented; for a single "judgment rule slice" case corpus that does not contain complete judgment rule structure elements, the legal terms are first preliminarily marked, and in combination with the legal terminology concept map, the relationship between the legal terms in the current judgment rule slice and the related legal terms in the case context and the attribution of the legal terms in the judgment rule structure model are identified, and based on this, the judgment rule structure type recognition is implemented.

[0018] In addition, the present invention also provides a judgment rule knowledge extraction system based on case text, the system comprising: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the judgment rule knowledge extraction method based on case text as described above.

[0019] After adopting such a design, the present invention has at least the following advantages:

[0020] The present invention is a method for constructing a rule map with case texts as the knowledge source. It mainly targets the judgment documents of criminal, civil and administrative first-instance, second-instance and retrial cases heard by the court. According to the business content characteristics, judgment document production specifications, semi-structured characteristics, composition characteristics and type characteristics of the judgment documents, an information model of the judgment rule map is designed, and a method for constructing legal terms, a classification method for identifying rules and a method for identifying rule types are proposed. Finally, the construction of the judgment rule map is realized, and an XML structured document is output, so that the computer system can effectively use the judgment rules existing in the case text to carry out legislation, law enforcement and judicial applications, and play a role in the fields of judgment document error correction, judgment document quality review, case quality supervision, and judicial case handling law and regulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Figure 1 is a flow chart of the method;

[0023] Figure 2 It is a schematic diagram of the principle of the slicing model;

[0024] Figure 3 It is a flowchart of building a legal terminology database;

[0025] Figure 4 It is a flowchart of candidate legal term generation;

[0026] Figure 5 It is a flowchart for constructing a concept map of legal terms. DETAILED DESCRIPTION

[0027] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0028] The present invention is a method for constructing a rule map with case texts as the knowledge source. It mainly targets the judgment documents of criminal, civil and administrative first-instance, second-instance and retrial cases heard by the court. According to the business content characteristics, judgment document production specifications, semi-structured characteristics, composition characteristics and type characteristics of the judgment documents, an information model of the judgment rule map is designed, and a method for constructing legal terms, a classification method for identifying rules and a method for identifying rule types are proposed. Finally, the construction of the judgment rule map is realized, and an XML structured document is output, so that the computer system can effectively use the judgment rules existing in the case text to carry out legislation, law enforcement and judicial applications, and play a role in the fields of judgment document error correction, judgment document quality review, case quality supervision, and judicial case handling law and regulations.

[0029] See also Figure 1 , the present invention comprises the following steps:

[0030] Step (1) According to the writing standards and structure of the judicial documents and the business standards of the litigation procedures, the documents are divided into multi-level text slices from coarse to fine, and a document slicing model is designed;

[0031] Step (2) based on the possible slicing positions of the referee rules, and based on the results of the text layered slicing in step (1), a hybrid algorithm model based on rules and deep learning is adopted to realize the design and algorithm implementation of the slicing model of the natural language segment where the referee rules are located;

[0032] Step (3) designing a structural model of the refereeing rules according to the business characteristics of the refereeing rules;

[0033] Step (4) Based on the own legal terminology database, the legal terminology database is constructed by using expert word database, rule-based bootstrapping method and mixed mode for the legal text corpus and judicial document text corpus;

[0034] Step (5) using the constructed legal terminology database and a deep learning-based relation extraction model to construct a legal terminology concept graph, the graph adopts standard triple storage;

[0035] Step (6) for the judgment rule slice corpus identified as “judgment rule”, according to the different types of features in the judgment rule model, the legal terminology library is integrated, and a hybrid method based on rules and deep learning is adopted to identify the relationship between the legal terms in the “judgment rule slice”, and according to the judgment rule structure model, the judgment rules existing in the judgment rule slice corpus are generated, and the link relationship between the judgment rules and the corpus is recorded;

[0036] Step (7) stores the extracted legal rules in the corresponding legal rule structure model according to the type of the extracted legal rules.

[0037] Step (1) According to the writing standards and structure of the judicial documents and the business standards of the litigation procedures, the documents are divided into multi-level text slices from coarse to fine, and a document slicing model is designed;

[0038] Judgment documents are usually written in WORD or WPS software, and the documents are clearly segmented, written in a standard manner, and have an obvious document structure. According to the business standards of the litigation procedure and summarizing its writing rules, this patent adopts a method of structuring judicial documents from coarse to fine. According to the court's requirements for document writing standards, the document is divided into multiple text pieces, and a document slicing model is designed to store each logical paragraph. The specific steps are as follows:

[0039] Step (1.1) summarizes the writing standards and structure of judicial documents, divides each paragraph of the document into multi-level text segments according to the logical relationship, and analyzes the first-level text segments of the judicial documents, including "text header", "basic information segment of the parties", "case origin segment", "litigation and defense information segment", "case facts segment", "judgment analysis process segment", "judgment result segment" and "text tail".

[0040] Step (1.2) Design a document slicing model to store each logical segment of the document. Each logical segment contains several fine slices. According to the fine slices contained in each paragraph, design a document slicing model, such as Figure 2 As shown, each fine slice content is stored in a string type, and the text slice is named according to the content it contains, such as "case fact segment" and "evidence fine slice segment". The entire slicing model is stored in a tree structure.

[0041] Step (2) based on the possible slicing positions of the referee rules, and based on the results of the text layered slicing in step (1), a hybrid algorithm model based on rules and deep learning is adopted to realize the design and algorithm implementation of the slicing model of the natural language segment where the referee rules are located;

[0042] Step (2.1) summarizes the target slice locations where the judgment rules may appear, pointing to the "case fact reasoning section", "evidence authentication section", "fact dispute focus identification section", "judgment summary section", and "legal dispute focus identification section".

[0043] Step (2.2) uses a rule-based NLP algorithm model based on the business connotation, location characteristics, and semantic characteristics of the judgment rules to extract the judgment rule slices in the target slice paragraphs of the case text.

[0044] Step (2.3) stores the extracted judgment rule slice text as a "judgment rule slice" node in the case xml overall model, and simultaneously records the original text location of the slice.

[0045] Step (3) Design the structural model of the arbitration rules by combining the constituent elements of the arbitration rules, the mandatory attributes of the arbitration rules, and the normative content. The specific steps are as follows:

[0046] Step (3.1) designs a structural model of standard arbitration rules, including three groups of elements: behavioral conditions, behavioral patterns, and legal consequences. Each group of elements is divided into two types: single factor and multi-factor. Two types of storage are supported in the structural model, and the logical types between factors in the case of multiple factors are designed; the elements of behavioral conditions, behavioral patterns, and legal consequences include two types: qualitative elements and quantitative elements. Quantitative elements support two modes: quantitative value and quantitative value range. Among the three groups of elements in the structural model of the arbitration rules, one of the behavioral conditions and behavioral patterns can be omitted.

[0047] Step (3.2) designs a structural model of non-standard adjudication rules. A certain behavior condition or behavior pattern can also serve as a legal consequence element in the above structural model. Other behavior conditions and behavior patterns serve as two groups of elements in the legal rule structural model and constitute a complete adjudication rule with this certain behavior condition or behavior pattern.

[0048] Step (3.3) designs the mandatory attributes of the refereeing rules, including three categories: imperative, prohibitive, and guiding.

[0049] Step (3.4) designs the normative content attributes of the arbitration rules, including authorization, obligation, and a combination of rights and obligations.

[0050] Step (4) Based on the own legal terminology database, a hybrid model of expert vocabulary, rule-based NLP and convolutional neural network is used to construct a legal terminology database for case text corpus. The method is as follows: Figure 3 .

[0051] The specific steps are as follows:

[0052] Step (4.1) prepares the legal text corpus and case text corpus.

[0053] The regulatory text corpus is processed through the following steps and provided in XML format:

[0054] (4.1.1) According to the writing standards and text structure of laws and regulations, and in accordance with legislative technical specifications, the legal text is segmented into multi-level text slices from coarse to fine, and a slicing model for the legal text is designed. Based on this slicing model, combined with the legislative technical features and semantic features of the legal text, a rule-based NLP algorithm is used to achieve slicing and storage of the legal text;

[0055] (4.1.2) Based on the legal principles, legislative technical specifications and application objectives of laws and regulations, and taking the entire text of the regulations as the object, design a regulatory attribute model; based on the results of text stratification and slicing in step (4.1.1), adopt a hybrid algorithm model based on rules and deep learning to extract and store regulatory attributes;

[0056] (4.1.3) Based on the legislative technical specifications and legal article attribute characteristics of the regulations, a deep learning algorithm model is used to achieve automatic classification of legal articles. The classification includes at least "legal concept articles", "legal principle articles", "legal rule articles", and "legislative purpose articles", and the article classification is stored in the "technical classification" attribute of the legal article.

[0057] The case text corpus is processed by steps (1) and (2) of this patent and provided in XML format.

[0058] Step (4.2) candidate legal terms are generated, such as Figure 4 shown.

[0059] Step (4.3) Calculate statistical features, calculate the TF-IDF, NC-Value, PMI, left (right) neighbor word entropy and other values ​​of the corpus as statistical features;

[0060] Step (4.4) classifier learning: learn a binary classifier based on positive and negative samples. The classifier model can be a decision tree, random forest or support vector machine.

[0061] Step (4.5) quality scoring, performing statistical feature calculation results;

[0062] Step (4.6) uses a convolutional neural network and the previously formed standard library for training to generate new legal terminology extraction results.

[0063] Step (5) constructs a legal terminology concept map. Based on step (4), a hybrid method based on pattern matching and deep learning is used to classify the relationships and store them in the form of triples to form a legal terminology concept map. The method is as follows: Figure 5 shown.

[0064] Step (5.1) designs a relationship type system in the legal terminology concept map, including synonym relationships, abbreviation relationships, hypernym relationships, hyponym relationships, reference relationships, and antonym relationships.

[0065] Step (5.2) is extracted based on pattern matching and bootstrapping.

[0066] Step (5.3) Expert engineering review to form standard data.

[0067] Step (5.4) Deep learning method, algorithm training.

[0068] Step (5.5) forms a legal terminology concept map based on the legal terminology database.

[0069] Step (6) for the judgment rule slice corpus identified as “judgment rule”, according to the different types of features in the judgment rule model, the legal terminology library is integrated, and a hybrid method based on rules and deep learning is adopted to identify the relationship between the legal terms in the “judgment rule slice”, and according to the judgment rule structure model, the judgment rules existing in the judgment rule slice corpus are generated, and the link relationship between the judgment rules and the corpus is recorded, which specifically includes:

[0070] Step (6.1) marks the legal terms for a single "judgment rule slice" case corpus containing complete elements of the judgment rule structure, and identifies the relationship between the legal terms and the attribution of the legal terms in the judgment rule structure model in combination with the currently marked legal terms and their context, and based on this, realizes the judgment rule structure type recognition. Type recognition is realized through the type of words in the legal terminology library, or the rule matching pattern based on pattern recognition, and the output result is standard judgment rule or non-standard judgment rule, as well as the single and multi-factor characteristics of behavioral conditions and behavioral patterns, and the quantitative and qualitative characteristics of legal consequences. Combined with the relationship between the identified legal terms, the judgment rule is generated.

[0071] Step (6.2) For a single "judgment rule slice" case corpus that does not contain complete elements of the judgment rule structure, first preliminarily mark the legal terms, and combine the legal terminology concept map to identify the legal terms of the current judgment rule slice plus the relationship between the related legal terms in the case context and the attribution of the legal terms in the judgment rule structure model, and based on this, realize the judgment rule structure type recognition. Through the type of words in the legal terminology library, or the rule matching pattern based on pattern recognition, the type recognition is realized, and the output result is a standard judgment rule or a non-standard judgment rule, as well as the single and multi-factor characteristics of behavioral conditions and behavioral patterns, the quantitative and qualitative characteristics of legal consequences, combined with the relationship between the identified legal terms, to generate the judgment rule.

[0072] Step (7) stores the extracted legal rules in the corresponding legal rule structure model according to the type of the extracted legal rules.

[0073] The following describes exemplary processing circuits of client, server, and cloud-based processing system resources for implementing algorithms and execution methods. A distributed processing system may include multiple instances of circuits that may be used to implement any processing circuit to perform Figure 1 , Figure 3 , Figure 4 and Figure 5The algorithm represented by the flowchart shown in . Not all components need to be used in various embodiments. For example, each of the clients, servers, and network resources of the distributed processing system can use a different set of components, and in the case of a graph database server, for example, a larger storage device can be used.

[0074] An exemplary processing system in the form of a computer may include a processing unit, a memory, a removable storage device, and a non-removable storage device, all coupled to a bus. The processing unit may include one or more single-core or multi-core processing devices. Although the exemplary processing system is described as a computer, the processing system may have different forms in different embodiments. For example, the processing system of the user device may also be a laptop, a tablet computer, or another processing device including an element that is the same or similar to the element. Devices such as laptops, tablet computers, etc. may be collectively referred to as mobile devices or user devices. In addition, although various data storage elements are described as part of a computer, the storage device may or may alternatively include a network connection (e.g., cloud-based) storage device that can be accessed through a network (e.g., a local area network (LAN), a personal area network (PAN), a wide area network (WAN) such as the Internet), or a local server-based storage device.

[0075] The memory may include volatile memory and non-volatile memory. The computer may include or have access to a processing environment including various computer-readable media (e.g., volatile memory, non-volatile memory, removable storage devices, and non-removable storage devices). Computer memory includes random access memory (RAM), read only memory (ROM), erasable programmable read-only memory (EPROM) and electrically erasable programmable read-only memory (EEPROM), flash memory or other storage technology, compact disc read-only memory (CDROM), digital versatile disk (DVD) or other optical disk storage, cassettes, magnetic tape, disk storage or other magnetic storage devices, or any other medium capable of storing computer-readable instructions.

[0076] The computer may include or have access to a processing environment including an input interface, an output interface, and a communication connection or interface, wherein the input interface, the output interface, and the communication connection or interface are connected to a bus. The output interface may include a display device, such as a touch screen or a computer display, and the output interface may also be used as an input device coupled to the input interface. The input interface may include one or more of a touch screen, a touch pad, a mouse, a keyboard, a camera, one or more device-specific buttons, one or more sensors integrated in the computer or coupled to the computer via a wired or wireless data connection, other input devices, etc. The computer may be operated in a network environment and connected to one or more remote computers, such as a mainframe, a server, and / or a database server, through a communication connection, which remote computers can be used to implement network connection services. User devices may include personal computers (PCs), servers, routers, network PCs, peer devices, or other common network nodes, etc. The communication connection may include a local area network (LAN), a wide area network (WAN), a cellular network, a Wi-Fi network, a Bluetooth network, the Internet, or other networks.

[0077] Computer readable instructions stored in a computer readable medium can be executed by a processing unit of a computer. Hard disk drives, CD-ROMs, and RAM are some examples of items that include non-transitory computer readable media (e.g., magnetic storage media, optical storage media, flash memory media, and solid-state storage media). Since carrier waves are too short-lived, the terms "computer readable medium" and "storage device" do not include carrier waves. For example, a processing unit can be caused to execute one or more methods or algorithms described herein through one or more application programs.

[0078] It should be understood that the software can be installed in one or more processors of user devices and / or network-connected services and sold with them. Alternatively, the software can be obtained through a physical medium or distributed system, such as from a server owned by the software creator or from a server not owned by the software creator but used by the software creator, and loaded into the user device and / or network-connected service. For example, the software can be stored on a server for distribution over the Internet.

[0079] Functions or algorithms described herein can be implemented by software in an embodiment.Software can include computer executable instructions stored in a computer-readable medium or a computer-readable storage device such as one or more physical storage devices or other types of hardware-based storage devices, local or on a network.In addition, these functions correspond to modules, and these modules can be software, hardware, firmware or other arbitrary combinations.Multiple functions can be performed in one or more modules as needed, and the described embodiments are merely exemplary.The software can be performed on a processing system such as a digital signal processor, an application-specific integrated circuit (ASIC), a microprocessor, a large processor or other types of processors running on a computer system (such as a personal computer server or other processing system), so as to change such processing systems into a specifically programmed machine.

[0080] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Those skilled in the art may make some simple modifications, equivalent changes or modifications using the technical contents disclosed above, which all fall within the protection scope of the present invention.

Claims

1. A method for extracting knowledge of adjudication rules based on case text, characterized in that: include: According to the writing standards and structure of judicial documents and the business standards of litigation procedures, the documents are divided into multi-level text slices from coarse to fine, and the document slicing model is designed; According to the possible slicing positions of the referee rules, based on the results of text layered slicing, a hybrid algorithm model based on rules and deep learning is adopted to realize the design and algorithm implementation of the slicing model of the natural language segment where the referee rules are located, including: Summarize the target slice positions that may appear in the adjudication rules, pointing to the "case fact reasoning section", "evidence authentication section", "fact dispute focus identification section", "adjudication essence section", and "legal dispute focus identification section"; Based on the business connotation, location characteristics, and semantic characteristics of the adjudication rules, a rule-based NLP algorithm model is used to extract the adjudication rule slices in the target slice paragraphs of the case text; The extracted judgment rule slice text is stored as a "judgment rule slice" node in the overall case xml model, and the original text location of the slice is recorded at the same time; According to the business characteristics of the referee rules, the referee rules structure model is designed, including: Combining the constituent elements of the adjudication rules, the mandatory attributes of the adjudication rules, and the normative content, the structural model of the adjudication rules is designed, including: Design a structural model of standard adjudication rules, including three groups of elements: behavior conditions, behavior patterns, and legal consequences. Each group of elements is divided into two types: single factor and multi-factor. The structural model supports two types of storage, and designs the logical types between factors in the case of multiple factors. The elements of behavior conditions, behavior patterns, and legal consequences include two types: qualitative elements and quantitative elements. Quantitative elements support two modes: quantitative value and quantitative value interval. Among the three groups of elements in the structural model of adjudication rules, one of the behavior conditions and behavior patterns can be omitted; Design a structural model of non-standard adjudication rules, where a certain behavior condition or behavior pattern is used as the legal consequence element in the above structural model, and other behavior conditions and behavior patterns are used as two groups of elements in the legal rule structural model to form a complete adjudication rule together with the certain behavior condition or behavior pattern; Design the mandatory attributes of the adjudication rules, including imperative, prohibitive, and guiding; The normative content attributes of designing adjudication rules include authorization, obligation, and combination of rights and obligations; Based on the proprietary legal terminology database, we use a hybrid model of expert word database, rule-based bootstrapping method and convolutional neural network to build a legal terminology database for the text corpus of regulations and judicial documents. Using the constructed legal terminology database and the deep learning-based relation extraction model, we construct a legal terminology concept map, which uses standard triple storage. For the judgment rule slice corpus identified as "judgment rules", according to the different types of features in the judgment rule model, the legal terminology library is integrated, and a hybrid method based on rules and deep learning is adopted to identify the relationship between legal terms in the "judgment rule slice", and according to the judgment rule structure model, the judgment rules existing in the judgment rule slice corpus are generated, and the link relationship between the judgment rules and the corpus is recorded, including: For a single "judgment rule slice" case corpus containing complete elements of the judgment rule structure, legal terms are marked, and the relationship between legal terms and their attribution in the judgment rule structure model are identified based on the currently marked legal terms and their context, and the judgment rule structure type is identified based on this; For a single "judgment rule slice" case corpus that does not contain complete elements of the judgment rule structure, we first preliminarily mark the legal terms, and then combine the legal term concept map to identify the legal terms of the current judgment rule slice and the relationship between the related legal terms in the case context and the attribution of the legal terms in the judgment rule structure model, and based on this, we can achieve judgment rule structure type recognition; According to the type of extracted legal rules, they are stored in the corresponding legal rule structure model.

2. The method for extracting knowledge of adjudication rules based on case text according to claim 1 is characterized in that: According to the writing standards and structure of judicial documents and the business standards of litigation procedures, the documents are divided into multi-level text slices from coarse to fine, and the document slicing model is designed, including: Summarize the writing standards and structure of judicial documents, divide the paragraphs of the documents into multi-level text segments according to the logical relationship, and analyze the first-level text segments of the judgment documents, including "text header", "basic information segment of the parties", "origin of the case segment", "pleading information segment", "case facts segment", "judgment analysis process segment", "judgment result segment" and "text tail"; A document slicing model is designed to store logical segments of documents. Each logical segment contains several fine slices. Based on the fine slices contained in each paragraph, a document slicing model is designed. The content of each fine slice is stored in a string type, and the text slice is named according to the content it contains. The names include: "case fact segment" and "evidence fine slice segment". The entire slicing model is stored in a tree structure.

3. The method for extracting knowledge of adjudication rules based on case text according to claim 1 is characterized in that: Based on the proprietary legal terminology database, we use a hybrid model of expert word database, rule-based bootstrapping method and convolutional neural network to build a legal terminology database for the text corpus of regulations and judicial documents, including: Prepare legal text corpus and case text corpus; Candidate legal term generation; Statistical feature calculation: TF-IDF, NC-Value, PMI, left neighbor word entropy and right neighbor word entropy of the corpus are calculated as statistical features; Classifier learning: Based on positive and negative samples, a binary classifier is learned. The classifier model uses a decision tree, random forest, or support vector machine. Quality scoring, performing statistical feature calculation results; A convolutional neural network is used to train the previously formed standard library to produce new legal terminology extraction results.

4. The method for extracting knowledge of adjudication rules based on case text according to claim 3 is characterized in that: Prepare legal text corpus and case text corpus, including: According to the writing standards and text structure of laws and regulations, and according to legislative technical specifications, the legal text is divided into multi-level text slices from coarse to fine, and a slicing model of the legal text is designed. Based on this slicing model, combined with the legislative technical features and semantic features of the legal text, a rule-based NLP algorithm is used to achieve slicing and storage of the legal text; According to the legal principles, legislative technical specifications and application objectives of laws and regulations, the entire regulatory text is taken as the object to design a regulatory attribute model. Based on the results of text layering and slicing, a hybrid algorithm model based on rules and deep learning is adopted to extract and store regulatory attributes. Based on the legislative technical specifications of laws and regulations and the attribute characteristics of legal provisions, a deep learning algorithm model is used to realize the automatic classification of legal provisions. The classification includes "legal concept provisions", "legal principle provisions", "legal rule provisions", and "legislative purpose provisions", and the provision classification is stored in the "technical classification" attribute of the legal provisions.

5. The method for extracting knowledge of adjudication rules based on case text according to claim 1 is characterized in that: Using the constructed legal terminology database and the deep learning-based relation extraction model, we construct a legal terminology concept map. The map uses standard triple storage, including: Design a relationship type system in the legal terminology concept map, including synonym relationships, abbreviation relationships, hypernym relationships, hyponym relationships, reference relationships, and antonym relationships; Extraction is performed based on pattern matching and bootstrapping; Expert engineering review to form standard data; Deep learning methods, algorithm training; A legal terminology concept map is formed based on the legal terminology database.

6. A judgment rule knowledge extraction system based on case text, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting knowledge of judgment rules based on case text according to any one of claims 1 to 5.

Images