A judgment document text extraction method and a computer readable storage medium

Abstract

The application discloses a judgment document text extraction method and a computer readable storage medium. The application is aimed at a judgment document information extraction task, optimizes the prior art, changes the input and output forms of downstream tasks into specified forms in a pre-training task, thereby unifying various information extraction tasks, and researchers do not need to construct different models for different tasks, and each extraction task does not need to be labeled separately. In this way, resource consumption can be greatly saved, the information extraction efficiency is improved, and the accuracy is relatively high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a judgment document text extraction method and a computer readable storage medium. BACKGROUND

[0002] The judgment document is a written conclusion with legal binding force made by a judge after the termination of a case trial, based on the facts of the case and legal provisions, on the entity and procedural issues of the case. A judgment document with complete structure, complete elements and rigorous logic is not only a certificate for the rights and obligations of the parties, but also an important basis for supervising civil trial activities.

[0003] The judgment document data formed in the judicial stage is non-structured text, and in the subsequent classification process, these non-structured texts need to be processed manually, which is very low in efficiency. Using BERT, Bi-LSTM, CRF and other technologies for document information extraction, the extraction process is relatively complex, so how to simply, conveniently and efficiently extract document information is a problem to be solved. SUMMARY

[0004] The present application provides a judgment document text extraction method and a computer readable storage medium to solve the problem that the prior art cannot simply, conveniently and efficiently extract document information.

[0005] In a first aspect, the present application provides a judgment document text extraction method, which comprises:

[0006] The judgment document to be extracted is screened to obtain a judgment document text meeting the preset classification requirements, and the reasons and facts in the screened judgment document are identified and extracted based on ontology design;

[0007] According to the received user indication, the target information in the judgment document to be obtained is determined, the ontology design is constructed, and the corresponding text content of the judgment document is extracted by using the constructed ontology design. The judgment document is trained based on the ontology design to obtain an extraction model, so that the trained extraction model can identify the information in the ontology design. The target information is the task of the judgment document.

[0008] The classified judgment document is extracted based on the trained extraction model, and the extraction model is further optimized by using the classified judgment document and the corresponding ontology design.

[0009] The ontology design is set according to the classification of the judgment document.

[0010] Optionally, the screening processing is performed on the to-be-extracted judicial documents to obtain judicial document texts meeting preset classification requirements, including: screening the to-be-extracted judicial documents to filter out repeated judicial documents and filter out judicial documents lacking the ontology design, and obtaining judicial document texts meeting preset classification requirements.

[0011] Optionally, the reason and fact part in the screened judicial documents is identified and extracted based on the ontology design, including: identifying whether the reason and fact part in the screened judicial documents includes the ontology design, and if yes, extracting the reason and fact part; and identifying all the screened judicial documents according to the identification step by step to obtain all the judicial documents including the ontology design from the screened judicial documents.

[0012] Optionally, the text content of the judicial documents is extracted based on the constructed ontology design, including: training the judicial documents based on the ontology design to obtain an extraction model.

[0013] The extracted text content of the judicial documents and the corresponding task are input into the extraction model with prompt words, and the task includes an entity recognition task, a relationship extraction task, an event extraction task and a sentiment extraction task.

[0014] The extraction model Wherein, s is a structured pattern director, s includes s1, s2,..., s j A plurality of structured pattern directors, x is a to-be-input sentence, x includes x1, x2,..., x k A plurality of sentences, M is an extraction model, y is an output of an extraction task, and j and k are natural numbers.

[0015]

[0016] Wherein, spot is a target name label in an information extraction task, and asso is a name label of a relationship between targets, and when the structured pattern director is extracted, spot and asso are added in front of the targets.

[0017] The structured pattern director s and the to-be-input sentence x are spliced and input into an encoder to obtain a hidden layer representation of each label token:

[0018] H = Encoder (s1,..., s |j| , x1,..., x |k| ), wherein s |j| is prompt information, and x |k| is a to-be-extracted sentence.

[0019] The target structured information is generated at the Decoder end using the hidden layer representation: y i ,h i d = Decoder([H; h1 d ,..., h i-1 d ]), y i is the i-th output information, h i d is the state of the i-th token, and H is the encoded information obtained in the previous step.

[0020] The input extraction task is converted into a preset extraction structure, that is, different extraction tasks are all divided into two extraction operations, wherein the first extraction operation is to find the target information segment in the input original sentence, and the second extraction operation is to find the relationship between the target information segments.

[0021] Optionally, the relationship between the target information segments includes a relationship between two entities in relation extraction or a relationship between an argument and a trigger word in event extraction.

[0022] Optionally, the loss function of the extraction model includes:

[0023] Capture the basic text-to-structure mapping capability: where θ e is the encoder parameter, θ d is the decoder parameter, (x, y) is the input sentence and the obtained output information pair, s meta is the meta-pattern combined with positive and negative examples, D pair is a set of binary tuples (x, y) output by each input x, and p is the conditional probability of y.

[0024] Loss function of the ability to generate effective structure: that is, taking the negative logarithm log of the probability p of each output result y, D recoder is a set of each information extraction record;

[0025] Loss function of semantic representation capability: D text is the set of sentences for each extraction task, x” is the masked text, x' is the source text, θ e is the encoder parameter, and θ d is the decoder parameter.

[0026] The three loss functions are added to obtain the final loss function: L = L Pair + L Record + L Text .

[0027] Optionally, the extracting and classifying the judgment document based on the trained extraction model comprises: performing entity extraction, relation extraction or event extraction on the judgment document to be classified by the trained extraction model, and outputting the extracted information to be output.

[0028] Optionally, the outputting the extracted information to be output comprises: outputting the extracted information in the form of a triple.

[0029] Optionally, the ontology design comprises one or more of a case type, a court of trial, a judgment date, a document type, a party, a case type, a case process, a case label, a legal basis and a document case type.

[0030] In a second aspect, the present application provides a computer readable storage medium, which stores a signal mapping computer program, and the computer program is executed by at least one processor to implement the text extraction method of the judgment document.

[0031] The present application has the following advantages:

[0032] The present application is aimed at the information extraction task of the judgment document, optimizes the prior art, changes the input and output form of the downstream task into a specified form in the pre-training task, thereby unifying various information extraction tasks, without the need for researchers to construct different models for different tasks, and without the need for separate annotation for each extraction task. In this way, resource consumption can be greatly saved, the efficiency of information extraction can be improved, and the accuracy can be ensured to be high.

[0033] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, and to implement the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0034] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present application. Moreover, the same reference numerals in the attached drawings indicate the same or similar components. In the drawings:

[0035] Figure 1 is a flowchart of a text extraction method of a judgment document provided by the first embodiment of the present application;

[0036] Figure 2 is a flowchart of a judgment document extraction process provided by the first embodiment of the present application;

[0037] Figure 3 is a flowchart of a training method of an extraction model provided by the first embodiment of the present application. DETAILED DESCRIPTION

[0038] The embodiments of the present application are aimed at the problems of low efficiency of artificial extraction and model extraction of judgment documents. By changing the input and output forms of downstream tasks into specified forms in pre-training tasks, the embodiments of the present application unify various information extraction tasks, without the need for researchers to build different models for different tasks or to separately annotate each extraction task. In this way, resource consumption can be greatly saved, the efficiency of information extraction can be improved, and the accuracy can be ensured. The present application will be further described in detail in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0039] Specifically, the embodiments of the present application are aimed at the methods of most current document information extraction, which use BERT, Bi-LSTM, CRF and other technologies. Although these technologies are effective, they are relatively complex in use. First, different models need to be built for different extraction tasks, which can result in low efficiency. For different application scenarios, researchers also need to design a large number of models for specific situations, which can consume a lot of resources. Second, the existing technology needs to annotate the text during information extraction, which is also a matter of consuming a lot of manpower and material resources. Because the extraction tasks are not unified, different extraction tasks need to be annotated separately, which consumes more resources. Third, the current full-supervised learning based on neural networks needs to combine feature learning with model training.

[0040] Based on the above problems, the embodiments of the present application propose a text extraction method for judgment documents. The method described in the embodiments of the present application can simplify various information extraction tasks while maintaining the accuracy of information extraction at a relatively high level, thereby ultimately improving the efficiency of information extraction. The method will be described in detail below in combination with Figures 1-3 The method described in the embodiments of the present application will be explained and described in detail through a specific example:

[0041] The first embodiment of the present application provides a text extraction method for judgment documents, referring to Figure 1 The method comprises:

[0042] S101, screening the judgment documents to be extracted to obtain judgment document texts meeting preset classification requirements;

[0043] That is, the embodiments of the present application screen the judgment documents to be extracted, thereby screening out repeated judgment documents and screening out judgment documents lacking a preset ontology design, and finally obtaining judgment document texts meeting preset classification requirements.

[0044] It should be noted that the preset ontology design of the embodiment of the present application can be a case, a court of trial, a judgment date, a document type, a party, a case type, a case process, a case label, a legal basis and a document case, etc. The person skilled in the art can set them arbitrarily according to actual needs, and the present application does not make specific limitations thereto.

[0045] In specific implementation, the embodiment of the present application is to collect original data of judgment documents, collect original data of judgment documents uploaded by users, and filter out original data that does not meet the requirements. In the filtering and screening, repeated judgment document data is removed through text similarity calculation, and data lacking content in the text, such as missing trial court and missing document case, is removed through keyword detection.

[0046] S102, identifying and extracting the reason and fact part in the screened judgment document based on the preset ontology design;

[0047] Specifically, the embodiment of the present application is to identify whether the reason and fact part in the screened judgment document includes the preset ontology design, and if so, to extract it.

[0048] According to the identification step, all screened judgment documents are identified to obtain all judgment documents including the preset ontology design in the screened judgment documents.

[0049] That is, the embodiment of the present application extracts the reason and fact part in the original data (i.e. the original judgment document) through trigger word identification, which is used for subsequent training model and information extraction.

[0050] In specific implementation, because the text of the judgment document is relatively standard, there will be a word indicating the information category before each kind of information, such as the words of “trial court”, “plaintiff”, “defendant”, etc. These words are trigger words, and after detecting the trigger words, the text content after the trigger words can be extracted. The specific extraction process of the judgment document is as shown in Figure 2

[0051] It should be noted that the ontology design in the embodiment of the present application is set according to the classification of the judgment document. That is, the person skilled in the art can set the ontology design according to the specific classification to meet different use requirements.

[0052] S103, determining the target information in the judgment document to be obtained according to the received user indication, constructing the ontology design, and extracting the corresponding text content of the judgment document by using the constructed ontology design, training the judgment document based on the ontology design to obtain an extraction model, so that the trained extraction model can identify the information in the ontology design;​

[0053] Among them, the target information of the embodiment of the application is various different tasks of judicial documents. In specific implementation, the embodiment of the application is to parse the received user instruction to obtain the required task of the judicial document, then construct the ontology design according to the required judicial document, and extract the text content corresponding to the judicial document by using the constructed ontology design, train the judicial document based on the ontology design to obtain the extraction model, so that the trained extraction model can identify the information in the ontology design;

[0054] Specifically, the embodiment of the application inputs the extracted text content of the judicial document and its corresponding task into the prompt word language extraction model, and the task includes entity recognition task, relationship extraction task, event extraction task and sentiment extraction task;

[0055] The extraction model Among them, s is a structured pattern director, s includes s1, s2,..., s |j| A plurality of structured pattern directors, x is a sentence to be input, x includes x1, x2,..., x |k| A plurality of sentences, M is a model, y is an output, and j and k are natural numbers;

[0056]

[0057] Among them, spot is the target name label in the information extraction task, and asso is the name label of the relationship between the targets. When using the structured pattern director to extract, spot and asso are added in front of the target;

[0058] The structured pattern director s and the sentence to be input x are spliced and input into the Encoder to obtain the hidden layer representation of each label token:

[0059] H=Encoder(s1,...,s |j| ,x1,...,x |k| ), wherein s |j| is a prompt information, and x |k| is a sentence to be extracted;

[0060] The hidden layer representation is used to generate target structured information at the Decoder end:

[0061] y i ,h i d =Decoder([H;h1 d ,...,h i-1 d ]),y i is the i-th output information, and hi d is the state of the i-th token, H is the encoding information obtained in the previous step;

[0062] The input extraction task is converted into a preset extraction structure, that is, different extraction tasks are all divided into two extraction operations, wherein the first extraction operation is to find the target information segment in the input original sentence, and the second extraction operation is to find the relationship between the target information segments.

[0063] The relationship between the target information segments includes the relationship between two entities in the relationship extraction or the relationship between the argument and the trigger word in the event extraction.

[0064] It should be noted that the trigger word in the embodiment of the application is used to extract case information in the structured judgment document, because the judgment document is written in a relatively standardized manner, and there are corresponding marker words before the information such as case number, case cause and processing result, so these words can be used as trigger words to extract the content after the trigger words; and the ontology design is the content that is artificially set to be extracted in the subsequent extraction task, which can be set to be the same as the trigger word or other case information different from the trigger word. The trigger word and the ontology design are used in different stages of the technology, the trigger word can only extract relatively standardized case information, and the ontology design can be designed to be more specific.

[0065] It should be noted that the loss function of the extraction model includes:

[0066] Capture the basic text-to-structure mapping capability: Wherein, θ e is the encoder parameter, θ d is the decoder parameter, (x, y) is the input sentence and the obtained output information pair, s meta is the meta-pattern combined with positive and negative examples, D pair is a set of binary tuples (x, y) output by each input x, and p is the conditional probability of y;

[0067] Loss function of the ability to generate effective structure: That is, the probability p of each output result y is taken as the negative logarithm log, D recoder is a set of each information extraction record;

[0068] Loss function of semantic representation capability: D text is the set of each extraction task sentence, x" is the masked text, x' is the source text, θ e is the encoder parameter, θ d is the decoder parameter;

[0069] The three loss functions are added to obtain a final loss function: L = L Pair + L Record + L Text .

[0070] In a specific implementation, the embodiment of the present application is designed by defining an ontology, that is, determining the information to be extracted, such as finding the cause of action, the court of trial, the date of judgment, the type of document, the parties, the type of case, the case process, the case label, the legal basis, and the cause of action of the document, which is used for subsequent training of the extraction model. Specifically, the embodiment of the present application is a model for the form of data to be added to the knowledge graph in a certain field by designing an ontology to standardize the expression form of structured data. In the embodiment of the present application, the ontology design is a set of information to be extracted for the judgment document data.

[0071] S104, based on the trained extraction model, the judgment document to be classified is extracted and classified, and the judgment document classified by the extraction model and the corresponding ontology design are used to further optimize the extraction model;

[0072] Specifically, the embodiment of the present application uses the trained model to realize the information extraction task of the judgment document based on the ontology design, such as entity extraction, relation extraction, event extraction, etc. Finally, the extracted information document is output according to the requirement, which can output the information to be extracted alone, such as the parties of each case, the judgment time, etc.; or output into a triple such as <entity, relation, entity>, <entity, attribute, entity>, which is used for the construction of the knowledge graph, etc. The specific model output form can be set arbitrarily according to the requirement, and the present application does not make a specific limitation.

[0073] Overall, the embodiment of the application is to use the prompt word paradigm to train the model, and the text content and the prompt words of different extraction tasks are input during the training of the model, so as to better extract the required information from the text. In addition, the method described in the embodiment of the application establishes a unified structure of the judgment document extraction task. Because the structures of different extraction tasks are generally different, the entity recognition task generally uses span and its entity category to represent, the relation extraction task generally uses a triplet structure to represent, the event extraction task generally uses a record to represent, and the opinion extraction task generally uses a triplet structure to represent. Using a unified information extraction structure, different extraction tasks can be modeled universally, the target structure can be generated adaptively, and the universal information extraction capability can be learned from different knowledge sources. Furthermore, unlike other extraction methods, the method described in the embodiment of the application can extract all the information required at one time, thereby saving resources for building different ontology designs.

[0074] In summary, the method described in the embodiment of the application is aimed at some judgment documents with long content and complex case, and it is a very energy-consuming thing for a judge to grasp the key information of the case and sort out the case in a short time. Most of the previous judgment document information extraction methods need to consume a large amount of resources to preprocess data, build ontology designs for different types of cases, and the extraction process is somewhat tedious and complex. The present application builds a unified information extraction structure to enable each information extraction task to be performed simultaneously and without repeated data labeling, and all the information required can be extracted at one time, including the information of the parties, time, process, and case. Therefore, the method described in the embodiment of the application greatly improves the efficiency in ontology design construction and text labeling, and prepares for subsequent information extraction, saving a large amount of resources and manpower, and ensuring that the accuracy of information extraction is not lower than that of the existing extraction methods.

[0075] The method described in the embodiment of the application will be described below with reference to a specific example. Figure 2 and Figure 3 The method described in the embodiment of the application will be described below with reference to a specific example.

[0076] S1, collect the original data of the judgment document, collect the original data of the judgment document uploaded by the user, and filter out the original data that does not meet the requirements. When filtering, remove the repeated judgment document data by text similarity calculation, and remove the data lacking content in the text by keyword detection, for example, the original data of the judgment document lacking the trial court and the document case can be filtered out;

[0077] S2, the reasons and facts in the original data are extracted by trigger word recognition for subsequent model training and information extraction;

[0078] Because the text of the judgment document is relatively standard, before each kind of information, there will be a word to explain the information category, such as "trial court", "plaintiff", "defendant" and other words, these words are trigger words, after detecting the trigger word, the text content after the trigger word can be extracted.

[0079] S3, self-defined ontology design, that is, to determine the information to be extracted, such as finding the case, the trial court, the judgment date, the type of document, the parties, the case type, the case process, the case label, the legal basis and the document case, these information is used for subsequent model training;

[0080] That is, the embodiment of the present application is to standardize the expression form of structured data by ontology design. In the present application, the ontology design is the collection of information to be extracted for the judgment document data;

[0081] S4, model training, first determine the types of tasks, such as entity recognition task, relationship extraction task, event extraction task and sentiment extraction task, here the prompt word paradigm is used to input the data required for training, the following is the principle of the model.

[0082] The text content and the prompt word corresponding to the task to be completed are input into the model, where s represents the structured mode guide, x represents the sentence to be input, M represents the model, and y represents the output:

[0083] Indicates that the sentence and the structured mode guide are input into the model

[0084]

[0085] (Calculation method of sentence and guide splicing)

[0086] =[[spot],...[spot]...,

[0087] [asso],...,[asso]...,

[0088] [text],x1,x2,...,x |x| ]

[0089] After splicing s and x, input into the Encoder, get the hidden layer representation of each label token:

[0090] H=Encoder(s1,...,s |s| ,x1,...,x |x| )

[0091] The spliced data is input into the encoder for encoding to obtain the state h of each token, wherein s is a prompt information, and x is a sentence to be extracted.

[0092] Then, the hidden layer representation is used to generate target structured information at the Decoder end:

[0093] y i ,h1 d = Decoder ([H; h1 d ,..., h i-1 d ])

[0094] In the decoder, the structured information y to be extracted is obtained.

[0095] The loss function of the pre-training task is given as follows:

[0096] L Pair represents a loss function for capturing the basic text-to-structure mapping capability:

[0097]

[0098] wherein θ e , θ d are parameters of the encoder and the decoder, (x, y) is a pair of input sentence and obtained output information, s meta is a meta-pattern combined by positive and negative examples;

[0099] L Record represents a loss function for generating valid structures:

[0100] L Text represents a loss function for semantic representation capability:

[0101] The three loss functions are added to obtain the final Loss: L = L Pair + L Record + L Text

[0102] Afterwards, the inputted each extraction task is converted into an extraction structure, that is, different extraction tasks are all split into two extraction operations, first, the target information segment in the inputted original sentence is found, such as a type of entity in entity recognition, trigger word and argument in event extraction are all segments in the original sentence, second, the relationship between the information segments outputted in the first step is found, such as the relationship between two entities in relation extraction, or the relationship between argument and trigger word in event extraction. Using the unified information extraction structure, different extraction tasks can be universally modeled, and the target structure can be adaptively generated, that is, a model capable of simultaneously performing entity recognition, relation extraction, event extraction and sentiment extraction can be trained.

[0103] S5, using the trained model based on the ontology design to perform information extraction tasks on the judicial documents, such as entity extraction, relation extraction, event extraction and the like, finally, according to the requirement, the extracted information document is outputted, the information to be extracted can be outputted separately, such as the parties of each case, the judgment time and the like, or the three tuples such as <entity, relation, entity>, <entity, attribute, entity> can be outputted, which can be used for the construction of a knowledge graph.

[0104] From the above, it can be known that the method described in the embodiment of the application is to obtain data from the judicial document network, and use the prompt word paradigm, unified extraction structure and the like, so that the process and difficulty of the information extraction task of the judicial document are greatly simplified, different extraction tasks are integrated, the workload required for labeling is reduced, and finally the information extraction is completed.

[0105] The second embodiment of the application provides a computer readable storage medium, the computer readable storage medium stores a signal mapping computer program, the computer program is executed by at least one processor, to realize the text extraction method of the judicial document in any one of the first embodiment of the application.

[0106] The related content of the embodiment of the application can be understood by referring to the first embodiment of the application, and will not be discussed in detail here.

[0107] Although the preferred embodiments of the application have been disclosed for the purpose of example, those skilled in the art will realize that various improvements, additions and substitutions are also possible, therefore, the scope of the application should not be limited to the above-mentioned embodiments.

Claims

1. A method for extracting text from court judgments, characterized in that, include: The selected judgment documents are screened to obtain judgment document texts that meet the preset classification requirements. The reasons and facts in the screened judgment documents are identified and extracted. Based on the received user instructions, the target information in the judgment document to be obtained is determined, an ontology design is constructed, and the constructed ontology design is used to extract the text content corresponding to the judgment document. The judgment document is trained based on the ontology design to obtain an extraction model, so that the trained extraction model can recognize the information in the ontology design. The target information is the task of the judgment document. Based on the extraction model obtained from training, the judgment documents to be classified are extracted and classified, and the extraction model is further optimized using the judgment documents classified by the extraction model and the corresponding ontology design. The ontology design is based on the classification of the judgment documents; The process of extracting text content corresponding to judgment documents using the constructed ontology design, and training an extraction model on the judgment documents based on the ontology design, includes: The extracted text content of the judgment document and its corresponding tasks are input into the extraction model using prompt word language. The tasks include entity recognition, relationship extraction, event extraction and sentiment extraction. The extraction model Where s is the structured pattern guide, and s includes Multiple structured pattern guides, For the sentence to be entered, The following includes Multiple sentences, M is the extraction model, y is the output of the extraction task, and j and k are both natural numbers; = Where spot is the name label of the target in the information extraction task, and ass is the name label of the relationship between the targets. When using the structured pattern guide for extraction, add spot and ass before the target. Connect the structured pattern instructors s and the sentence to be input. After concatenation, the data is input into the encoder to re-encode s and x, obtaining the hidden representation of each tag token: ,in It's a notification message. The sentences to be extracted; Using hidden layer representations to generate target structured information at the decoder end: , For the i-th output information, Let H represent the state of the i-th token, and H be the encoded information obtained in the previous step. The input extraction tasks are transformed into a preset extraction structure, that is, each extraction task is split into two extraction operations. The first extraction operation is to find the target information fragment in the original input sentence, and the second extraction operation is to find the relationship between the target information fragments. The relationships between the target information fragments include the relationship between two entities in relation extraction or the relationship between arguments and trigger words in event extraction. The loss function of the extraction model includes: Capture basic text-to-structure mapping capabilities: ,in, These are encoder parameters. These are the decoder parameters, where (x, y) is the input sentence and the output information pair. The meta-pattern is a combination of positive and negative samples. The set of pairs (x, y) that output y for each input x. The conditional probability of y; The loss function for the ability to generate effective structures: That is, taking the negative logarithm (log) of the probability p for each output result y. A collection of records extracted for each piece of information; Loss function for semantic representation capability: , A collection of sentences for each extraction task. For the concealed text, For the source text, These are encoder parameters. These are decoder parameters; The final loss function is obtained by adding the three loss functions together: .

2. The method according to claim 1, characterized in that, The process of filtering the extracted judgment documents to obtain judgment document texts that meet the preset classification requirements includes: The extracted judgment documents are screened to remove duplicate judgment documents and judgment documents that lack the ontology design, so as to obtain judgment document texts that meet the preset classification requirements.

3. The method according to claim 1, characterized in that, The process of identifying and extracting the reasoning and factual portions of the filtered court documents based on ontology design includes: Identify whether the reasoning and factual sections of the filtered judgment documents include ontology design; if so, extract it. The process involves identifying all the filtered judgment documents one by one according to the identification steps, in order to extract all judgment documents containing ontology design from the filtered judgment documents.

4. The method according to any one of claims 1-3, characterized in that, The extraction model, trained from scratch, extracts and classifies the judgment documents to be classified, including: The extraction model obtained through training performs entity extraction, relation extraction, or event extraction on the judgment documents to be classified, and outputs the extracted information.

5. The method according to claim 4, characterized in that, The step of outputting the extracted information includes: The extracted information is output in the form of triples.

6. The method according to any one of claims 1-3, characterized in that, The ontology design includes one or more of the following: cause of action, court of trial, date of judgment, document type, parties, case type, case process, case tag, legal basis, and document cause of action.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that maps signals, which, when executed by at least one processor, implements the text extraction method for judicial documents as described in any one of claims 1-6.

Images