CN115878815A - Legal document judgment result prediction method, device and storage medium - Google Patents

Abstract

The present invention relates to the technical field of data processing, and relates to a judgment result prediction method, device and storage medium of legal documents. The method includes: obtaining legal judgment documents of massive legal judgment examples; extracting legal elements and legal The relationship between the elements is determined to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result; according to the judgment event chain and element relationship corresponding to each legal judgment document, the legal judgment instance is centered on the construction of legal The knowledge map of the judgment document; receiving the input legal facts, wherein the legal facts include legal claims; using the knowledge map and the legal retrieval model to predict and output the legal judgment results corresponding to the legal facts. Through this technical solution, the complex and redundant work of aligning and mapping complex and redundant elements to inference functions is omitted, and good results are achieved in document retrieval based on excellent feature representation.

Description

Judgment result prediction method, device and storage medium for legal documents

technical field

The present invention relates to the technical field of data processing, in particular to a method, device and storage medium for predicting a judgment result of a legal document.

Background technique

In recent years, big data and artificial intelligence algorithms have been increasingly valued and promoted in various industries and fields. Many countries, including my country, have raised artificial intelligence to a national strategic level. In the judicial field, benefiting from the promotion of big data technology, judicial institutions at all levels in my country have entered a period of "smart court" construction with the provision of smart judicial services as the core.

However, the task-specific approaches proposed in the current field of smart justice research still have fundamental obstacles to their practical application. One is that the mainstream methods at this stage are mostly based on models such as machine learning and neural networks. The black box defects of the models lead to the general lack of interpretability of the research process and results, which greatly reduces the credibility and usability of the models. The second is that models that rely on large-scale data training generally lack reasoning mechanisms. Machine intelligence generally refers to the ability of an agent to learn, perceive, understand, and work like humans. Understanding the characteristics of human cognition is one of the necessary conditions for realizing intelligence. First, knowledge reasoning is an important means of human cognition. Most methods based on statistical models today cannot use knowledge reasoning to obtain results, that is, the model does not have a reasoning mechanism.

Judicial reasoning is a method for courts to obtain judgments in cases and an important means to prove the legitimacy of judicial decisions. Therefore, it is not only a method of legal thinking, but also a practical rationality or practical reasoning process for judges to solve problems. Theoretically speaking, judicial decisions should be the logical result of judicial reasoning. In the judgment of any case in a society ruled by law, judges should provide certain reasons or grounds, and judicial reasoning can provide justification for judicial decisions, because the primary role of legal reasoning is to provide justification for conclusions. At the same time, a logic Strict judicial reasoning itself forms a strong reason or basis.

The existing reasoning logic mainly has the following defects:

1) Some application logic generally has a large number of legal elements that need to be aligned to standard categories, such as child custody judgments in divorce judgments 'child is less than 2 years old' needs to be aligned to a large number of reasons such as "child|age|less than 2 years old" Alignment, if the condition is missing, it cannot be reasoned, which is not an easy task for knowledge system construction and batch reasoning.

2) Usually the existing reasoning mode is to construct a decision function, which is basically rule-based reasoning, and does not have the diversity, fault tolerance and early warning functions of big data.

3) The existing element extraction analysis model is too dependent on large batches of data, and has no reliable robustness and small-sample learning ability.

Contents of the invention

In order to overcome the problems existing in related technologies, the present invention provides a judgment result prediction method, device and storage medium of legal documents, thereby reducing the labor cost of obtaining data, improving domain adaptability, and better adapting to fine-grained applications Scenes.

According to a first aspect of an embodiment of the present invention, a method for predicting a judgment result of a legal document is provided, the method comprising:

Obtain legal judgment documents of a large number of legal judgment examples;

Extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, wherein the legal elements include the plaintiff's claim type, Judgment reasons and judgment results, the element relationship includes causality and correspondence;

According to the judgment event chain and element relationship corresponding to each legal judgment document, the knowledge graph of the legal judgment document is constructed centering on the legal judgment instance, wherein the knowledge graph includes from top to bottom: root node, cause of action category, appeal category , legal judgment examples, judgment reasons and judgment results, and the knowledge map also includes annotations of element relationships;

receiving input legal facts, wherein the legal facts include legal claims;

Using the knowledge graph and the legal retrieval model, predict and output the legal judgment results corresponding to the legal facts.

In one embodiment, preferably, the extraction of legal elements and the determination of the relationship between legal elements are performed on each legal judgment document, including:

Using the element category classification model to identify the legal elements of each of the legal judgment documents to obtain the element category classification and identification results;

Splicing the recognition result of the element category classification and the features of the legal judgment document, and adding case category information into the reading comprehension model, so that the reading comprehension model determines the content and location of the legal element according to the element category classification recognition result , output the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result.

In one embodiment, preferably, the method further includes:

Using Milvus as the feature memory, an index set is built from top to bottom according to the knowledge map, and the legal coding model of the transformer is used to encode the legal judgment document features, document vector features, judgment event chains, event chain vectors, and legal judgments The document ID is correspondingly stored.

In one embodiment, preferably, using the knowledge graph and the legal retrieval model, predict and output the legal judgment results corresponding to the legal facts, including:

Perform intent analysis on the legal facts to determine the target cause category and target appeal category corresponding to the legal facts;

Searching in the knowledge map according to the target case category and the target appeal category to determine at least one corresponding legal judgment instance;

Using the transformer legal coding model to perform feature coding on the legal facts and the corresponding judgment reasons of at least one legal judgment instance, to obtain the corresponding legal fact feature vector and the judgment reason feature vector of the legal judgment instance;

Calculate the similarity between the legal fact feature vector and the judgment reason feature vector of the legal judgment instance, and determine the legal judgment result corresponding to the legal fact according to the similarity.

In one embodiment, preferably, the method further includes:

receiving an input search command for a target legal fact;

According to the retrieval command, using the transformer legal coding model to perform feature coding on the target legal facts to obtain the coded features;

Perform similarity matching between the encoded features and the stored features of the legal judgment documents and the chain of judgment events, so as to retrieve the judgment documents corresponding to the target legal facts.

In one embodiment, preferably, the method further includes:

Receiving an input similar document search command;

According to the similar document retrieval command, determine the judgment event chain corresponding to the current legal document;

Use the Transformer legal coding model to encode the features of the judgment event chain of the current legal document, and obtain the encoded event chain features;

Similarity matching is performed on the coded event chain features with the stored legal judgment document features and judgment event chains, so as to retrieve legal judgment documents similar to the current legal documents.

In one embodiment, preferably, the method further includes:

The transformer model is learned and trained according to the legal judgment documents of a large number of legal judgment examples, so as to obtain the transformer legal coding model.

According to a second aspect of an embodiment of the present invention, a device for predicting a judgment result of a legal document is provided, and the device includes:

The obtaining module is used to obtain legal judgment documents of a large number of legal judgment instances;

The first determination module is used to extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, wherein the The legal elements include the type of plaintiff's claim, the reason for the judgment and the result of the judgment, and the relationship between the elements includes causality and correspondence;

The building module is used to construct the knowledge map of the legal judgment document centering on the legal judgment instance according to the judgment event chain and element relationship corresponding to each legal judgment document, wherein the knowledge map includes from top to bottom: root node, Category of cause of action, category of appeal, instance of legal judgment, reason for judgment and result of judgment, and the knowledge map also includes annotation of element relationship;

a receiving module, configured to receive input legal facts, wherein the legal facts include legal claims;

The prediction module is configured to use the knowledge graph and the legal retrieval model to predict and output the legal judgment results corresponding to the legal facts.

In one embodiment, preferably, the first determination module is used for:

Using the element category classification model to identify the legal elements of each of the legal judgment documents to obtain the element category classification and identification results;

Splicing the recognition result of the element category classification and the features of the legal judgment document, and adding case category information into the reading comprehension model, so that the reading comprehension model determines the content and location of the legal element according to the element category classification recognition result , output the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result.

In one embodiment, preferably, the device further includes:

The storage module is used to use Milvus as the feature memory, construct an index set according to the top-down hierarchy of the knowledge map, and use the transformer legal coding model to encode the legal judgment document features, document vector features, judgment event chains, and event chain vectors , and correspondingly stored with the legal judgment document ID.

In one embodiment, preferably, the prediction module is used for:

Perform intent analysis on the legal facts to determine the target cause category and target appeal category corresponding to the legal facts;

Searching in the knowledge map according to the target case category and the target appeal category to determine at least one corresponding legal judgment instance;

Using the transformer legal coding model to perform feature coding on the legal facts and the corresponding judgment reasons of at least one legal judgment instance, to obtain the corresponding legal fact feature vector and the judgment reason feature vector of the legal judgment instance;

Calculate the similarity between the legal fact feature vector and the judgment reason feature vector of the legal judgment instance, and determine the legal judgment result corresponding to the legal fact according to the similarity.

In one embodiment, preferably, the device further includes:

A first retrieval module, configured to receive an input retrieval command for a target legal fact;

The first encoding module is used to perform feature encoding on the target legal facts according to the retrieval command using the Transformer legal encoding model to obtain encoded features;

The first matching module is configured to perform similarity matching between the encoded features and the stored features and judgment event chains of the legal judgment documents, so as to retrieve the judgment documents corresponding to the target legal facts.

In one embodiment, preferably, the device further includes:

The second retrieval module is used to receive an input similar document retrieval command;

The second determination module is used to determine the judgment event chain corresponding to the current legal document according to the similar document retrieval command;

The second encoding module is used to use the transformer legal encoding model to encode the features of the judgment event chain of the current legal document, and obtain the encoded event chain features;

The second matching module is configured to perform similarity matching between the coded event chain features and the stored legal judgment document features and judgment event chains, so as to retrieve legal judgment documents similar to the current legal documents.

In one embodiment, preferably, the device further includes:

The training module is used to learn and train the transformer model according to the legal judgment documents of a large number of legal judgment examples, so as to obtain the transformer legal coding model.

According to a third aspect of an embodiment of the present invention, a device for predicting a judgment result of a legal document is provided, the device comprising:

processor;

memory for storing processor-executable instructions;

Wherein, the processor is configured as:

Obtain legal judgment documents of a large number of legal judgment examples;

Extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, wherein the legal elements include the plaintiff's claim type, Judgment reasons and judgment results, the element relationship includes causality and correspondence;

According to the judgment event chain and element relationship corresponding to each legal judgment document, the knowledge graph of the legal judgment document is constructed centering on the legal judgment instance, wherein the knowledge graph includes from top to bottom: root node, cause of action category, appeal category , legal judgment examples, judgment reasons and judgment results, and the knowledge map also includes annotations of element relationships;

receiving input legal facts, wherein the legal facts include legal claims;

Using the knowledge graph and the legal retrieval model, predict and output the legal judgment results corresponding to the legal facts.

According to a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the method described in any one of the embodiments of the second aspect is implemented. step.

The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

In the embodiment of the present invention, the causal relationship of chain structure is constructed, from the cause of action to the category of appeal to the reason of judgment to the judgment result, and a deeper analysis and expression are constructed and associated one by one to build a legal judgment knowledge map, through the customized semantic model Table answering simplifies multi-round reasoning functions into semantic understanding tasks, omits the complex and redundant work of aligning and mapping complex and redundant elements to reasoning functions, and achieves good results in document retrieval based on excellent feature table answers.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description serve to explain the principles of the invention.

Fig. 1 is a flowchart of a method for predicting a judgment result of a legal document according to an exemplary embodiment.

Fig. 2 is a schematic composition diagram of a legal judgment document according to an exemplary embodiment.

Fig. 3 is a schematic diagram of a decision event chain according to an exemplary embodiment.

Fig. 4 is a schematic diagram of a knowledge graph according to an exemplary embodiment.

Fig. 5 is a flow chart of step S102 in a method for predicting a judgment result of a legal document according to an exemplary embodiment.

Fig. 6A is a schematic diagram of a specific model structure and processing process of an element category classification model and a reading comprehension model according to an exemplary embodiment.

Fig. 6B is a schematic diagram showing label naming according to an exemplary embodiment.

Fig. 7 is a schematic diagram of feature storage according to an exemplary embodiment.

Fig. 8 is a flow chart of step S105 in a method for predicting a judgment result of a legal document according to an exemplary embodiment.

Fig. 9 is a flow chart showing another method for predicting a judgment result of a legal document according to an exemplary embodiment.

Fig. 10 is a flowchart showing yet another method for predicting a judgment result of a legal document according to an exemplary embodiment.

Fig. 11 is a block diagram of an apparatus for predicting a judgment result of a legal document according to an exemplary embodiment.

Detailed ways

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with aspects of the invention as recited in the appended claims.

Fig. 1 is a flowchart of a method for predicting a judgment result of a legal document according to an exemplary embodiment.

As shown in FIG. 1 , according to the first aspect of the embodiments of the present invention, a method for predicting a judgment result of a legal document is provided, and the method includes:

Step S101, obtaining legal judgment documents of a large number of legal judgment instances;

The legal judgment documents mainly include six main parts: party information, trial process, plaintiff’s statement, defendant’s defense, court investigation, and this court’s opinion, as well as modeling from three perspectives (plaintiff’s perspective, defendant’s perspective, and court’s perspective), as shown in Figure 2 As shown, among them, legal instrument indicates the entire judgment document, party information indicates the party information of the entire judgment document, trial process indicates the trial process of the entire judgment document, plaintiff claim indicates the plaintiff claim of the entire judgment document, and defender defense The defendant who means the entire judgment document argues, court investigation means the court’s investigation of the entire judgment document, and court believes means the opinion of the entire judgment document.

There are also main perspectives, among which plaintiff view represents the perspective of the plaintiff, defender view represents the perspective of the defendant, and court view represents the perspective of the court.

Step S102, extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, wherein the legal elements include plaintiff The type of appeal, the reason for the judgment and the result of the judgment, and the relationship between the elements includes the causal relationship and the corresponding relationship;

Since the present invention mainly constructs the causal relationship of the judgment, that is, the reasoning of the court, the court's perspective part is mainly refined here, which can be modeled as the plaintiff's claim type (judgment type), judgment reason (condition), and judgment result:

Among them, court believes means the content of the paragraph that the court believes, and the analysis of this paragraph is divided into multiple elements, among which claim is the type of plaintiff's appeal, reason means the reason (condition) of the judgment, and res means the result of the judgment.

Judgment causality extraction of legal documents can be regarded as the structural extraction of elements of the court’s reasoning part, which is mainly aimed at element extraction and causality correspondence for the court’s reasoning part of paragraphs or lengths, including: plaintiff’s claim type (claim), judgment The reason (reason), judgment result (res), and the corresponding relationship are shown in Figure 3, presenting a (1, n, 1) one-to-many-to-1 relationship. Finally, <claim, r1, reason> (where r1 represents the corresponding relationship), <reason, r2, res> (where r2 represents the causal relationship) is represented in a chain structure.

Traditional methods of relation extraction are generally defined as two types of tasks, namely text extraction and relation prediction (graph neural network). However, due to the problem of data cost in the real generation environment, the data for the decision type is sparse and the overall energy can be supplied to the model. Due to various reasons such as the amount of data and the complexity of relational labeling, the learning efficiency of traditional relational analysis models cannot reach the expected level. Therefore, the present invention proposes a method based on data association and progressive relation extraction to analyze the chain structure from appeal type to decision reason (condition) and finally to decision result.

Step S103, according to the judgment event chain and element relationship corresponding to each legal judgment document, construct the knowledge map of the legal judgment document centering on the legal judgment instance, wherein the knowledge map includes from top to bottom: root node, cause of action category , appeal category, legal judgment instance, reason for judgment and judgment result, and the knowledge map also includes annotation of element relationship;

The knowledge graph is shown in Figure 4. The root node is the core node of the entire graph; the substructure is the casecategory; the lower level of the case category is the claim category, which includes the core common demands under the cause of the case. This section uses the divorce judgment as an example. For example: Divorce Judgment --> Custody Judgment | Divorce Property Judgment. A many-to-one or one-to-one relationship between the decision reason and the decision res is constructed centering on each decision instance example, and each subgraph represents an instance of the appeal category. Continuously accumulate the causal relationship of judgments to meet the reasoning needs of judgments.

Step S104, receiving input legal facts, wherein the legal facts include legal claims;

Step S105, using the knowledge graph and the legal retrieval model to predict and output the legal judgment result corresponding to the legal fact.

In the embodiment of the present invention, the causal relationship of chain structure is constructed, from the cause of action to the category of appeal to the reason of judgment to the judgment result, and a deeper analysis and expression are constructed and associated one by one to build a legal judgment knowledge map, through the customized semantic model Table answering simplifies multi-round reasoning functions into semantic understanding tasks, omits the complex and redundant work of aligning and mapping complex and redundant elements to reasoning functions, and achieves good results in document retrieval based on excellent feature table answers.

Fig. 5 is a flow chart of step S102 in a method for predicting a judgment result of a legal document according to an exemplary embodiment.

As shown in Figure 5, in one embodiment, preferably, step S102 includes:

Step S501, using the element category classification model to identify the legal elements of each of the legal judgment documents, so as to obtain the element category classification recognition result;

Step S502, splicing the identification result of the element category classification and the features of the legal judgment document, and adding the case category information into the reading comprehension model, so that the reading comprehension model can determine the legal element according to the element category classification identification result Content and location, output the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result.

Overcome the key problem of less information in small sample datasets through the efficient learning efficiency of classification prediction. Then the results of the classification prediction part can clearly indicate the element category information in this sentence. The reading comprehension model (mrc) based on the mechanism of prompt learning reduces the learning cost of the reading comprehension model (mrc) and greatly improves the reading comprehension model (mrc) during prediction. )s efficiency.

Among them, the specific model structure and processing process of the element category classification model and the reading comprehension model are shown in Figure 6A. The high learning efficiency of classification prediction overcomes the key problem of less information in small sample data sets. Then the results of the classification prediction part can clearly indicate the element category information in this sentence. The reading comprehension model (mrc) based on the mechanism of prompt learning reduces the learning cost of the reading comprehension model (mrc) and greatly improves the reading comprehension model (mrc) during prediction. )s efficiency. Among them, input(sentence) represents the text input at the sentence level; classer represents the classification model of the element categories contained in the sentence; mrc input represents the input splicing of reading comprehension, including the result of the classifier res1query as the splicing input(sentence) of reading comprehension The text sentence of the final is the overall feature of the mrc model; mrc means the reading comprehension model; output means that the model output includes the position category in the entire chain structure, such as {...} appeal type, {...} reason for judgment, {. ..} Judgment results and specific element information; C indicates that the category division of the cause of action provides preconditions for model learning to improve learning efficiency; element category task indicates the element category data contained in short sentences; mrc task indicates that the reading comprehension task is used to extract Detailed elements.

Specifically, as shown in Figure 6A, the feature category classification model includes:

Input (sentence) layer: The input layer is the original text input by the model. The legal judgment text is divided into individual small sentences according to symbols, and then the small sentences in a paragraph are input into the model in batches for model analysis. The category of the case (C) improves the model learning efficiency.

classer layer: The element category classifier mainly analyzes the element categories contained in the sentence, and the feature representation layer is also called the embedding layer, which uses the neural network to represent the category of the input text. Commonly used feature representation methods include the use of convolutional neural networks, long and short memory machines, and transformer series models. The present invention uses the lawformer model.

As shown in Figure 6B, element categories can be marked in the form of tags. Subsequently, the relationship labels are converted into a chain structure in the form of label naming, which facilitates the subsequent conversion of inefficient relationship learning models into efficient classification learning models.

Reading comprehension models include:

mrc input layer: The reading comprehension input mainly includes the recognition results of the three-party information classification model, the original legal text, and the suggested cause of action category (C). Added cause category (C) information.

mrc layer: mainly analyze the detailed content and location of elements based on the previous progressive information. The main technical means include name recognition (ner) and segment classification (span_class). The present invention adopts the reading comprehension model (mrc) to achieve the control purpose of extracting a certain type of element separately and greatly improves efficiency and reduces calculation amount.

output layer: output the model results, forming a chain structure including: {...} appeal type, {...} judgment reason, {...} judgment result and specific element information.

In one embodiment, preferably, the method further includes:

As shown in Figure 7, Milvus is used as the feature memory, and an index set is constructed according to the top-down hierarchy of the knowledge map, and the legal coding model of the transformer is used to encode the legal judgment document features, document vector features, judgment event chain, and event chain vector, and store it correspondingly with the ID of the legal judgment document.

As shown in Figure 8, in one embodiment, preferably, step S105 includes:

Step S801, performing intent analysis on the legal facts to determine the target cause category and target appeal category corresponding to the legal facts;

Step S802, searching in the knowledge graph according to the target case category and the target appeal category to determine at least one corresponding legal judgment instance;

Step S803, using the transformer legal coding model to perform feature encoding on the legal facts and the corresponding judgment reason of at least one legal judgment instance, and obtain the corresponding legal fact feature vector and the judgment reason feature vector of the legal judgment instance;

Step S804, calculating the similarity between the legal fact feature vector and the judgment reason feature vector of the legal judgment instance, and determining the legal judgment result corresponding to the legal fact according to the similarity.

The invention transforms the reasoning problem of legal judgment into the legal text retrieval problem. Traditional text matching processes the features at the character level or word level, and uses the model to learn the character or word features of the text. Since this method cannot fully learn the semantics of the text, the recall ability of legal text retrieval is not strong or the judgment events do not match, and the feature representation of the text is the prerequisite for improving legal text retrieval.

The present invention enhances the characteristics of the vertical field in the legal field based on the transformer model, which is mainly divided into two aspects: 1. Large batches of texts in the legal field; . On the basis of the original transformer model, the algorithm’s ability to understand legal texts and feature retrieval capabilities is improved. More importantly, the model can better adapt to characteristic scenarios such as party information, fact descriptions, court opinions, and judgment results in judgments, and has achieved Nice effect.

After receiving legal facts, text preprocessing can be carried out first, and conventional text denoising methods can be carried out, including: removing special symbols, removing redundant blanks, and converting traditional Chinese to simplified Chinese, etc., and then analyzing the intent of the legal text, including division of case, appeal category divided. According to the results of the analysis, the classification of the case and the classification of appeals reach the instance layer from the upper level of the map. Then use the customized Transformer legal retrieval model to perform feature representation. There are two main data sources here:

A. Judgment reason data: directly convert the judgment reason into transformer fixed-length features. If there are multiple reasons, splice the reasons into a paragraph and then convert it into transformer fixed-length features.

B. Judgment document data: analyze the court reasoning part of the judgment text, parse it into appeal categories, judgment reasons, and judgment results, and then express it into fixed-length features through the transformer.

Based on the reason for the decision, the data is converted into a fixed-length feature of the transformer representation, the most likely result is inferred based on the vector calculation module, and the bad result is discarded according to the threshold, so as to achieve the function of causal reasoning. Statistical reasoning methods can also be used to count the proportions of different judgment results in similar judgment situations, so as to achieve the purpose of analysis and early warning.

Among them, the judgment results mainly include:

(1) Whether the claim is supported (sales contract disputes - the validity of the clauses: the commercial housing sales contract relationship is established and legal and valid; sales contract disputes - whether the claims are valid: the claims are valid).

(2) Judge the plaintiff or the defendant (divorce judgment-custody: awarded to the plaintiff).

As shown in Figure 9, in one embodiment, preferably, the method further includes:

Step S901, receiving an input search command for the target legal fact;

Step S902, according to the search command, use the Transformer legal coding model to perform feature coding on the target legal fact to obtain the coded features;

Step S903, performing similarity matching on the encoded feature with the stored feature of the legal judgment document and the judgment event chain, so as to retrieve the judgment document corresponding to the target legal fact.

As shown in Figure 10, in one embodiment, preferably, the method further includes:

Step S1001, receiving an input similar document retrieval command;

Step S1002, according to the similar document retrieval command, determine the judgment event chain corresponding to the current legal document;

Step S1003, using the transformer legal coding model to perform feature coding on the judgment event chain of the current legal document to obtain the coded event chain features;

Step S1004, performing similarity matching between the coded event chain features and the stored legal judgment document features and judgment event chains, so as to retrieve legal judgment documents similar to the current legal documents.

Focus on the core paragraphs of the judgment document (this court thinks) and construct the core logic of the judgment document, that is, the judgment logic, and construct the judgment logic into a chain structure, encode the chain structure into core features based on the customized legal version transformer, and perform vector Calculation, the main features are hierarchical analysis categories, overall document features, core judgment logic features, and combined calculation scores to retrieve similar documents.

In one embodiment, preferably, the method further includes:

The transformer model is learned and trained according to the legal judgment documents of a large number of legal judgment examples, so as to obtain the transformer legal coding model.

Fig. 11 is a block diagram of an apparatus for predicting a judgment result of a legal document according to an exemplary embodiment.

As shown in FIG. 11, according to the second aspect of the embodiment of the present invention, a device for predicting a judgment result of a legal document is provided, and the device includes:

Obtaining module 1101, configured to obtain legal judgment documents of a large number of legal judgment instances;

The first determination module 1102 is used to extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, where all The aforementioned legal elements include the type of plaintiff's claim, the reason for the judgment and the result of the judgment, and the relationship between the aforementioned elements includes causal relationship and corresponding relationship;

The construction module 1103 is used to construct the knowledge graph of the legal judgment document centering on the legal judgment instance according to the judgment event chain and element relationship corresponding to each legal judgment document, wherein the knowledge graph includes from top to bottom: a root node , category of cause of action, category of appeal, legal judgment instance, reason for judgment and judgment result, and the knowledge map also includes annotation of element relationship;

A receiving module 1104, configured to receive input legal facts, wherein the legal facts include legal claims;

The prediction module 1105 is configured to use the knowledge graph and the legal retrieval model to predict and output the legal judgment results corresponding to the legal facts.

In one embodiment, preferably, the first determination module is used for:

Using the element category classification model to identify the legal elements of each of the legal judgment documents to obtain the element category classification and identification results;

Splicing the recognition result of the element category classification and the features of the legal judgment document, and adding case category information into the reading comprehension model, so that the reading comprehension model determines the content and location of the legal element according to the element category classification recognition result , output the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result.

In one embodiment, preferably, the device further includes:

The storage module is used to use Milvus as the feature memory, construct an index set according to the top-down hierarchy of the knowledge map, and use the transformer legal coding model to encode the legal judgment document features, document vector features, judgment event chains, and event chain vectors , and correspondingly stored with the legal judgment document ID.

In one embodiment, preferably, the prediction module is used for:

Perform intent analysis on the legal facts to determine the target cause category and target appeal category corresponding to the legal facts;

Searching in the knowledge map according to the target case category and the target appeal category to determine at least one corresponding legal judgment instance;

Using the transformer legal coding model to perform feature coding on the legal facts and the corresponding judgment reasons of at least one legal judgment instance, to obtain the corresponding legal fact feature vector and the judgment reason feature vector of the legal judgment instance;

Calculate the similarity between the legal fact feature vector and the judgment reason feature vector of the legal judgment instance, and determine the legal judgment result corresponding to the legal fact according to the similarity.

In one embodiment, preferably, the device further includes:

A first retrieval module, configured to receive an input retrieval command for a target legal fact;

The first encoding module is used to perform feature encoding on the target legal facts according to the retrieval command using the Transformer legal encoding model to obtain encoded features;

The first matching module is configured to perform similarity matching between the encoded features and the stored features and judgment event chains of the legal judgment documents, so as to retrieve the judgment documents corresponding to the target legal facts.

In one embodiment, preferably, the device further includes:

The second retrieval module is used to receive an input similar document retrieval command;

The second determination module is used to determine the judgment event chain corresponding to the current legal document according to the similar document retrieval command;

The second encoding module is used to use the transformer legal encoding model to encode the features of the judgment event chain of the current legal document, and obtain the encoded event chain features;

The second matching module is configured to perform similarity matching between the coded event chain features and the stored legal judgment document features and judgment event chains, so as to retrieve legal judgment documents similar to the current legal documents.

In one embodiment, preferably, the device further includes:

The training module is used to learn and train the transformer model according to the legal judgment documents of a large number of legal judgment examples, so as to obtain the transformer legal coding model.

According to a third aspect of an embodiment of the present invention, a device for predicting a judgment result of a legal document is provided, the device comprising:

processor;

memory for storing processor-executable instructions;

Wherein, the processor is configured as:

Obtain legal judgment documents of a large number of legal judgment examples;

Extract legal elements and determine the relationship between legal elements for each legal judgment document, so as to obtain the judgment event chain and element relationship from the plaintiff's claim type, judgment reason to judgment result, wherein the legal elements include the plaintiff's claim type, Judgment reasons and judgment results, the element relationship includes causality and correspondence;

According to the judgment event chain and element relationship corresponding to each legal judgment document, the knowledge graph of the legal judgment document is constructed centering on the legal judgment instance, wherein the knowledge graph includes from top to bottom: root node, cause of action category, appeal category , legal judgment examples, judgment reasons and judgment results, and the knowledge map also includes annotations of element relationships;

receiving input legal facts, wherein the legal facts include legal claims;

Using the knowledge graph and the legal retrieval model, predict and output the legal judgment results corresponding to the legal facts.

According to a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the method described in any one of the embodiments of the second aspect is implemented. step.

It can be further understood that "plurality" in the present invention refers to two or more, and other quantifiers are similar. "And/or" describes the association relationship of associated objects, indicating that there may be three types of relationships, for example, A and/or B may indicate: A exists alone, A and B exist simultaneously, and B exists independently. The character "/" generally indicates that the contextual objects are an "or" relationship. The singular forms "a", "said" and "the" are also intended to include the plural unless the context clearly dictates otherwise.

It can be further understood that the terms "first", "second", etc. are used to describe various information, but the information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another, and do not imply a specific order or degree of importance. In fact, expressions such as "first" and "second" can be used interchangeably. For example, without departing from the scope of the present invention, first information may also be called second information, and similarly, second information may also be called first information.

It can be further understood that although operations are described in a specific order in the drawings in the embodiments of the present invention, it should not be understood as requiring that these operations be performed in the specific order shown or in a serial order, or that Perform all operations shown to obtain the desired result. In certain circumstances, multitasking and parallel processing may be advantageous.

Other embodiments of the invention will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present invention, these modifications, uses or adaptations follow the general principles of the present invention and include common knowledge or conventional technical means in the technical field not disclosed in the present invention . The specification and examples are to be considered exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

It should be understood that the present invention is not limited to the precise constructions which have been described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims (10)

1. A method for predicting a decision result of a legal instrument, the method comprising:

obtaining a legal decision document of a mass legal decision instance;

extracting legal elements and determining the relationship between the legal elements for each legal decision document to obtain a decision event chain and an element relationship from an original appeal type, a decision reason to a decision result, wherein the legal elements comprise the original appeal type, the decision reason and the decision result, and the element relationship comprises a causal relationship and a corresponding relationship;

according to a judgment event chain and an element relation corresponding to each legal judgment document, constructing a knowledge graph of the legal judgment documents by taking a legal judgment example as a center, wherein the knowledge graph sequentially comprises from top to bottom: the knowledge graph also comprises element relation labels;

receiving an input legal fact, wherein the legal fact comprises a legal appeal;

and predicting and outputting a legal judgment result corresponding to the legal fact by using the knowledge graph and the legal retrieval model.

2. The method of claim 1, wherein the extracting of legal elements and the determining of the relationship between the legal elements for each legal decision document comprises:

carrying out legal element identification on each legal decision document by using an element type classification model to obtain an element type classification identification result;

and splicing the element category classification and identification result and the characteristics of the legal judgment document, adding case category information into the reading understanding model, inputting the case category information into the reading understanding model, so that the reading understanding model determines the content and the position of the legal element according to the element category classification and identification result, and outputting a judgment event chain and an element relation from the original report request type, the judgment reason to the judgment result.

3. The method of claim 1, further comprising:

and constructing an index set according to the top-down hierarchy of the knowledge graph by taking Milvus as a feature storage, coding by using a transformer legal coding model to obtain legal decision document features, document vector features, decision event chains and event chain vectors, and correspondingly storing the legal decision document ID.

4. The method of claim 1, wherein predicting and outputting legal decision results corresponding to the legal facts by using the knowledge graph and a legal retrieval model comprises:

performing intention analysis on the legal fact to determine a target case category and a target appeal category corresponding to the legal fact;

retrieving in the knowledge graph according to the target case routing category and the target appeal category to determine at least one corresponding legal decision instance;

performing feature coding on the legal fact and the judgment reason of at least one corresponding legal judgment instance by using a transform legal coding model to obtain a corresponding legal fact feature vector and a judgment reason feature vector of the legal judgment instance;

and calculating the similarity of the legal fact feature vector and the judgment reason feature vector of the legal judgment example, and determining a legal judgment result corresponding to the legal fact according to the similarity.

5. The method of claim 3, further comprising:

receiving an input retrieval command for the target legal fact;

according to the retrieval command, performing feature coding on the target legal fact by using a transformer legal coding model to obtain coded features;

and performing similarity matching on the coded features and the stored legal decision document features and decision event chains to retrieve a decision document corresponding to the target legal fact.

6. The method of claim 3, further comprising:

receiving an input similar document retrieval command;

determining a judgment event chain corresponding to the current legal document according to the similar document retrieval command;

performing feature coding on a judgment event chain of the current legal document by using a transformer legal coding model to obtain coded event chain features;

and performing similarity matching on the coded event chain features and the stored legal decision document features and decision event chains to retrieve a legal decision document similar to the current legal document.

7. The method of claim 3, further comprising:

and learning and training the transform model according to the law decision documents of the massive law decision examples to obtain the transform legal coding model.

8. An apparatus for predicting a decision result of a legal document, comprising:

the system comprises an acquisition module, a judgment module and a judgment module, wherein the acquisition module is used for acquiring legal judgment documents of a large number of legal judgment instances;

the first determining module is used for extracting legal elements and determining the relationship between the legal elements for each legal judgment document to obtain a judgment event chain and an element relationship from an original report appeal type, a judgment reason to a judgment result, wherein the legal elements comprise the original report appeal type, the judgment reason and the judgment result, and the element relationship comprises a causal relationship and a corresponding relationship;

the construction module is used for constructing a knowledge graph of the legal decision documents by taking the legal decision examples as a center according to the decision event chain and the element relation corresponding to each legal decision document, wherein the knowledge graph sequentially comprises from top to bottom: the knowledge graph also comprises element relation labels;

the receiving module is used for receiving the input legal fact, wherein the legal fact comprises legal requirements;

and the prediction module is used for predicting and outputting a legal judgment result corresponding to the legal fact by using the knowledge graph and the legal retrieval model.

9. An apparatus for predicting a decision result of a legal document, comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to:

obtaining a legal decision document of a mass legal decision instance;

extracting legal elements and determining the relationship between the legal elements for each legal judgment document to obtain a judgment event chain and an element relationship from an original appeal type, a judgment reason to a judgment result, wherein the legal elements comprise the original appeal type, the judgment reason and the judgment result, and the element relationship comprises a causal relationship and a corresponding relationship;

according to a judgment event chain and a factor relation corresponding to each legal judgment document, constructing a knowledge graph of the legal judgment documents by taking a legal judgment example as a center, wherein the knowledge graph sequentially comprises from top to bottom: the knowledge graph also comprises element relation labels;

receiving an input legal fact, wherein the legal fact comprises a legal appeal;

and predicting and outputting a legal judgment result corresponding to the legal fact by using the knowledge graph and the legal retrieval model.

10. A computer-readable storage medium having stored thereon computer instructions, which, when executed by a processor, carry out the steps of the method according to any one of claims 1 to 7.

Images