Model training method, legal action information alignment and fusion method, and terminal device thereof

Abstract

The application provides a model training method, a multi-source heterogeneous legal litigation information alignment and fusion method, a terminal device and a computer readable storage medium. The method comprises the following steps: extracting litigation subjects and their association relationships in a plurality of multi-source heterogeneous data sources, and constructing litigation subject knowledge triples according to the litigation subjects and their association relationships; using the litigation subject knowledge triples corresponding to the same litigation subject to form a first knowledge graph, and using the litigation subject knowledge triples corresponding to different litigation subjects to form a second knowledge graph; pairing and combining the first knowledge graph and the second knowledge graph in a double circulation mode, and training a judgment model according to the feature vectors between the knowledge graphs. The application automatically generates a large amount of knowledge graph pairing training data by matching the knowledge graph pairs of the same and different subjects in a double circulation mode, and automatically sets a label for whether the same subject according to the pairing information, so as to provide sufficient data support for the training of the graph attention judgment model and reduce the demand for manual labeling.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the technical field of law and deep learning, and particularly relates to a model training method, a multi-source heterogeneous legal litigation information alignment and fusion method, a terminal device and a computer readable storage medium. BACKGROUND

[0002] The legal litigation process is complex and changeable. When a lawyer conducts case analysis, a large amount of multi-source heterogeneous data related to litigation subjects often needs to be analyzed, and the efficiency of manual data processing is low. Although current intelligent technologies such as natural language processing and knowledge graph can effectively improve the manual processing speed, the existing technologies can only process a single data source, extract litigation entity information from a single data source (for example, a legal document), and cannot perform subject information fusion and alignment between multiple data sources that may have different data structures (such as multiple unstructured document data and structured subject association information obtained from a commercial website), resulting in low matching accuracy and still requiring a large amount of manual checking and alignment work, which further reduces work efficiency and cannot effectively support further analysis and decision-making. Therefore, the existing legal document intelligent analysis technology needs to be improved, and an intelligent litigation subject information alignment and fusion method is urgently needed. SUMMARY

[0003] The present application provides a model training method, a multi-source heterogeneous legal litigation information alignment and fusion method, a terminal device and a computer readable storage medium, aiming to solve the problem in the prior art that legal analysis personnel need to manually check subject information involved in multi-source heterogeneous legal litigation, thereby reducing work efficiency.

[0004] To solve the above technical problems, the present application provides a model training method, which comprises:

[0005] extracting litigation subjects and their association relationships in a plurality of multi-source heterogeneous data sources, and constructing litigation subject knowledge triples according to the litigation subjects and their association relationships;

[0006] composing a first knowledge graph by using litigation subject knowledge triples corresponding to the same litigation subject, composing a second knowledge graph by using litigation subject knowledge triples corresponding to different litigation subjects, and composing a litigation subject training data set according to the first knowledge graph and the second knowledge graph;

[0007] performing vectorization representation on the knowledge graphs in the litigation subject training data set;

[0008] pairing and combining the first knowledge graph and the second knowledge graph in a double-loop manner, training a judgment model according to the feature vectors between the knowledge graphs, obtaining a judgment model for comparing the similarity of any litigation subject, and aligning the litigation subjects of the same subject.

[0009] wherein the extraction of the litigation subject and the associated relationship in the plurality of multi-source heterogeneous data sources comprises:

[0010] the litigation subject and the associated relationship in the plurality of multi-source heterogeneous data sources are extracted through a flexible deep learning information extraction method;

[0011] wherein the multi-source heterogeneous data source comprises structured data and unstructured data provided by different data sources, different data interfaces or access methods.

[0012] wherein the flexible deep learning information extraction method comprises: constructing a flexible information extraction tool that adapts to the data characteristics according to the type of the multi-source heterogeneous data source.

[0013] wherein the construction of the litigation subject knowledge triple according to the litigation subject and the associated relationship comprises:

[0014] the litigation subject is taken as a point of the litigation subject knowledge triple;

[0015] the associated relationship of the litigation subject is taken as an edge of the litigation subject knowledge triple;

[0016] a litigation subject knowledge triple is constructed by using two points and an edge connecting the points.

[0017] wherein the vectorization representation of the knowledge graph in the litigation subject training data set comprises:

[0018] obtaining a word pre-training vector dictionary in an open source legal litigation field text corpus;

[0019] according to the point and edge name semantics in the litigation subject training data set, searching in the vector dictionary to obtain a point and edge semantic vector, and realizing the vectorization representation of the knowledge graph;

[0020] using an attention model to integrate graph point, edge semantics and associated topological structure feature information on the litigation subject point, and generating a knowledge graph feature vector of the litigation subject.

[0021] wherein the pairing combination of the first knowledge graph and the second knowledge graph in a double cycle mode comprises:

[0022] cyclically traversing the knowledge graph of each litigation subject in the litigation subject training data set;

[0023] based on the current knowledge graph, pairing with other graphs of the same subject, and pairing with graphs in the data set of other different litigation subject knowledge graphs, automatically setting labels for the same graphs and different graphs of the subject;

[0024] In each pair of matched two knowledge graphs, the difference between the predicted label and the actual label is updated according to the judgment model, and the model training is completed.

[0025] Wherein, after training the judgment model according to the feature vectors between the knowledge graphs, obtaining the judgment model for comparing the similarity of any litigation subject, and aligning the litigation subjects of the same subject, the model training method comprises:

[0026] Obtain any pair of knowledge graphs, wherein the indication graph comprises a pair of litigation subjects and their corresponding association relationship;

[0027] Using the judgment model, generate the current pair of litigation subject knowledge graph vectors and calculate the similarity;

[0028] Determine whether the paired subjects are the same litigation subject by threshold, and fuse the alignment of the subjects considered as the same subject.

[0029] To solve the above technical problems, the present application provides a multi-source heterogeneous legal litigation information alignment fusion method, which comprises:

[0030] Obtain the target litigation subject type;

[0031] Obtain the subject association relationship information corresponding to the target litigation subject type;

[0032] Based on the target litigation subject type and its subject association relationship information, form a plurality of knowledge graphs with the target litigation subject type as the center point from a plurality of legal litigation information data sources;

[0033] Use a preselected trained judgment model to match the plurality of knowledge graphs and obtain the similarity between the plurality of knowledge graphs;

[0034] Align the litigation subjects of the plurality of knowledge graphs with a similarity exceeding a preset threshold as the same litigation subject, and realize the fusion of multi-source heterogeneous legal information;

[0035] Wherein, the judgment model is trained by the above model training method.

[0036] To solve the above technical problems, the present application provides a terminal device, wherein the terminal device comprises a processor and a memory connected to the processor, wherein the memory stores program instructions;

[0037] The processor is configured to execute the program instructions stored in the memory to implement the above model training method and / or multi-source heterogeneous legal litigation information alignment fusion method.

[0038] To address the aforementioned technical problems, this application provides a computer-readable storage medium storing program instructions that, when executed, implement the aforementioned model training method and / or the multi-source heterogeneous legal litigation information alignment and fusion method.

[0039] Compared with the prior art, the beneficial effects of this application are as follows: The terminal device extracts the litigation subjects and their relationships from several multi-source heterogeneous data sources, and constructs litigation subject knowledge triples according to the litigation subjects and their relationships; it uses litigation subject knowledge triples corresponding to the same litigation subject to form a first knowledge graph, and uses litigation subject knowledge triples corresponding to different litigation subjects to form a second knowledge graph, and forms a litigation subject training dataset based on the first knowledge graph and the second knowledge graph; it performs vectorization representation on the knowledge graphs in the litigation subject training dataset; it uses a double-loop method to pair and combine the first knowledge graph and the second knowledge graph, trains a judgment model based on the feature vectors between the knowledge graphs, and obtains a judgment model for comparing the similarity of any litigation subject and aligning the judgments of litigation subjects with the same subject. The method proposed in this application addresses the problem that existing automated legal information processing technologies cannot automatically compare and align subject information within multi-source heterogeneous data, resulting in the need for extensive manual verification, significantly reducing work efficiency and failing to effectively support further analysis and decision-making. It automatically generates a large amount of knowledge graph pairing training data by matching knowledge graph pairs of the same and different subjects through a dual-loop approach. It also automatically sets labels based on pairing information to indicate whether the subjects are the same, providing sufficient data support for training the graph attention judgment model and reducing the need for manual annotation. Furthermore, it extracts litigation subjects and their relationships from adapted multi-source heterogeneous data, overcoming the limitation of existing information extraction and natural language processing methods in the legal litigation field, which can only handle single data sources and struggle to effectively identify litigation subject information in multi-source heterogeneous data sources. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0041] Figure 1 This is a flowchart illustrating an embodiment of the model training method provided in this application;

[0042] Figure 2 This is a schematic diagram of the training process of the judgment model provided in this application;

[0043] Figure 3 This is a schematic diagram of the structure of an embodiment of the judgment model training set provided in this application;

[0044] Figure 4 This is a schematic diagram illustrating the working principle of the attention mechanism of the judgment model provided in this application;

[0045] Figure 5 This is a flowchart illustrating an embodiment of the multi-source heterogeneous legal litigation information alignment and fusion method provided in this application;

[0046] Figure 6 This is a schematic diagram of the framework of an embodiment of the terminal device provided in this application;

[0047] Figure 7 This is a schematic diagram of the structure of an embodiment of the computer storage medium provided in this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0049] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0050] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0051] Lawyers need to analyze whether the litigants described in large amounts of heterogeneous data from multiple sources are the same entity when conducting case analysis. Existing automated legal information analysis technologies can only process single data sources and cannot automatically compare and align subject information across heterogeneous data sources, still requiring a significant amount of manual verification work, thus reducing the efficiency of the analysis process. Therefore, existing intelligent legal document analysis technologies need improvement, and there is an urgent need for intelligent methods for aligning and integrating litigant information.

[0052] To address the aforementioned shortcomings of existing technologies, this application provides a method for aligning and fusing multi-source heterogeneous legal litigation information, as well as a method for training the judgment model used therein. This application solves the problem that existing automated legal information processing technologies cannot automatically compare subject information within multi-source heterogeneous data, resulting in the need for extensive manual verification, significantly reducing work efficiency, and failing to effectively support further analysis and decision-making.

[0053] Please refer to details. Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating an embodiment of the model training method provided in this application. Figure 2 This is a schematic diagram of the training process of the judgment model provided in this application.

[0054] like Figure 1 As shown, the model training method in this embodiment specifically includes the following steps:

[0055] Step S11: Extract the litigation subjects and their relationships from several heterogeneous multi-source data sources, and construct litigation subject knowledge triplets according to the litigation subjects and their relationships.

[0056] Specifically, the litigants in this example include natural persons and various types of entities. Entities include state-owned enterprises, private enterprises, public institutions, non-profit organizations, and other entities involved in socio-economic activities. Relationships between individuals include kinship, friendship, and professional / official relationships. Relationships between individuals and entities include equity relationships, employment relationships, and financial relationships (such as loan relationships and pledge / mortgage relationships). For attribute relationships, those related to individuals include address (e.g., residential address, office address), date (e.g., date of birth), gender, and age; those related to entities include legal person status, category (e.g., state-owned enterprise, private enterprise, public institution), address (e.g., former and current location), and date (e.g., registration date, capital increase date, listing date).

[0057] Terminal devices can extract litigation subjects and their relationships from several multi-source heterogeneous data sources using flexible deep learning information extraction methods; among them, multi-source heterogeneous data sources include structured data and unstructured data stored in different data sources and provided through different data interfaces or access methods.

[0058] Specifically, the multi-source nature of heterogeneous legal litigation data is reflected in the fact that the data is stored in different data sources and accessed through different interfaces, such as local computer documents, open-source websites (e.g., legal document websites), commercial websites (e.g., Qichacha, Aiqicha), and social media. The heterogeneity of this data lies in the fact that it includes both structured data, such as structured tabular data (e.g., enterprise information tables), and unstructured data, such as legal documents, website announcements, and social media comments. By inputting this heterogeneous data into a flexible information extraction model, the litigation subjects and their relationships are extracted. The model automatically matches the litigation subjects and relationships extracted from different data sources, generating several knowledge triples.

[0059] In one implementation, the litigation subjects and their relationships are extracted by a flexible information extraction tool, which includes regular expressions, logical rules, and deep learning models, and can be adjusted according to the characteristics of the legal litigation data.

[0060] Specifically, this application categorizes multi-source heterogeneous data into tabular standardized data and non-tabular non-standardized data based on the degree of structuring and the standardization of semantic expression. The terminal device of this application can automatically determine the standardization of input data, including using heuristic logic rules to determine the data format based on the hypertext tags of the input data, and then extracting the main information using the corresponding extraction method.

[0061] This includes standardized tabular data. For example, publicly available basic information about litigation entities from platforms like Qichacha has a standardized header-value data structure. This example extracts litigation entity information by combining logical rules and regular expressions. For instance, regular expressions can determine whether text at the current table position belongs to the header (e.g., "Company Name" or the value "A Certain Technology Company"). Logical rules, using an "if-then" approach, clean the table value text and extract key information based on features such as text length, subject-verb-object position, and order.

[0062] Non-tabular, non-standard data includes, but is not limited to, online announcements, social media comments, and various legal texts such as judgments and rulings. While non-tabular, non-standard data may have certain standard sentence structures, it lacks a standardized structure and exhibits significant randomness and variability due to manual writing. This example uses a deep learning model (Bi-LSTM-CRF) integrating bidirectional long short-term memory networks to extract the main information from unstructured legal litigation data.

[0063] The input to the Bi-LSTM-CRF model is a legal litigation text sentence, and the output is the words contained in the input sentence that express the subjects of the legal litigation and their relationships. For example, "a company" and "a person's name" express the subjects of the litigation, "an address" and "a telephone number" express the attributes of the subjects, and "employment," "friends," "parents," "mortgage," and "pledge" express the relationships between the subjects.

[0064] The Bi-LSTM-CRF model requires training based on manually labeled samples. These manually labeled samples are similar in format to the model's output samples, including accurate labels for the litigants and their related entities within legal text sentences, using the "Beginning-Middle-End-Other" (BIEO) principle (e.g., labeling all specific company names with the uniform "Company" label). Model training involves inputting the original text into the model, comparing the model's output with the manually labeled real samples, and updating the model parameters using cross-entropy loss, backpropagation, and gradient descent methods. The trained Bi-LSTM-CRF deep learning model can receive legal text and automatically extract words expressing the litigants and their relationships.

[0065] Furthermore, the terminal device constructs several knowledge graphs centered on different litigation subjects using the subjects and relationships extracted from the Bi-LSTM-CRF deep learning model in the form of "point (head entity)-edge-point (tail entity)" triples. In these triples, the head entity represents the litigation subject, the tail entity represents an attribute of another litigation subject or the current litigation subject, and the edges represent the relationships between subjects or between a subject and an attribute. This example designs logical reasoning rules to establish the litigation subject knowledge triples for both tabular and non-tabular data. For standardized tabular data, the logical rule set matches the table headers and values ​​based on the table's design distribution; for non-standardized non-tabular data, the logical rule set matches the extracted words representing the litigation subject and their associated positions.

[0066] For example, when the input data is a structured table, the logical reasoning rules can traverse all rows and columns of the table. If the cell to the right of the current header is a value, and a header appears again to the right of that value, then the value matches the previous header. When the input data is unstructured text, if the Bi-LSTM-CRF deep learning model extracts the labels "company" and "address," it assumes that there is a "subject-attribute" relationship between the specific company and address referred to by the text.

[0067] Step S12: Use the litigation subject knowledge triples corresponding to the same litigation subject to form a first knowledge graph, use the litigation subject knowledge triples corresponding to different litigation subjects to form a second knowledge graph, and use the first knowledge graph and the second knowledge graph to form a litigation subject training dataset.

[0068] Specifically, to achieve the alignment and fusion of litigation subjects within multi-source heterogeneous data, this application trains a judgment model. The judgment model is trained using a training set composed of knowledge graph datasets of multiple litigation subjects. The knowledge graph datasets include knowledge triples describing the same litigation subject from different data sources, forming the subject's knowledge graph dataset. The knowledge graphs are vectorized, including integrating graph feature information at the litigation subject's center point using pre-trained word vectors and a graph attention model (GAT). A dual-loop pairing method is used, including a first knowledge graph set of knowledge graphs for the same subject and a second knowledge graph set of knowledge graphs for different subjects. Knowledge graph pairings are sampled through iterative traversal, and labels are automatically set based on known pairing information to indicate whether the currently paired subjects are the same subject. The judgment model is trained using SoftMarginLoss loss based on the pairing labels.

[0069] Specifically, the terminal device constructs a knowledge graph dataset for a specific litigation subject. This includes manually cross-referencing and identifying triples describing the same subject extracted from different data sources. The triples from each data source automatically form a knowledge graph centered on the same litigation subject. In other words, multiple knowledge graphs exist centered on the same litigation subject depending on the data source, collectively forming the knowledge graph dataset for that subject.

[0070] In addition, the terminal device constructs a training set for the judgment model, including applying dataset construction methods to establish knowledge graph datasets of multiple different litigation subjects to form a training set for the judgment model. For example... Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of an embodiment of the judgment model training set provided in this application. The training set includes several different sets of knowledge graphs representing different litigation subjects. Each set of knowledge graphs includes several knowledge graphs representing the subject extracted from different data sources, and each knowledge graph includes several triples. The knowledge graph dataset and training set can be saved and visualized using the Neo4j graph database tool.

[0071] Step S13: Vectorize the knowledge graph within the training dataset of the litigation subjects.

[0072] Specifically, the terminal device initializes the semantic information of points and edges within the knowledge graph of each litigation subject using pre-trained word vectors, realizing vectorized representation of points and edges; and uses a graph attention model to integrate the semantic information features of points and edges of the current knowledge graph and topological structure information features at the central point of the litigation subject to generate a knowledge graph feature vector.

[0073] Specifically, the terminal device obtains a pre-trained word vector dictionary for the legal litigation domain from an open-source dataset, where each word corresponds to a semantic numerical vector of 50-300 dimensions. Based on the semantics of the point and edge names within the main knowledge graph, it searches the open-source word vector set to obtain the pre-trained vectors corresponding to their names, and initially vectorizes the names of all points and edges within the knowledge graph to express their semantic feature information.

[0074] The terminal device traverses each point in each subject's knowledge graph, collecting points and edges directly associated with the current point to form multiple triples c. i，j，p =(h i r p h j ), where h i and h j r is the semantic vector of the corresponding point. p This is the semantic vector for the corresponding edge. Then, several triples corresponding to each subject knowledge graph are input into the graph attention model to generate vectors that can simultaneously express the semantic and topological features of the graph's points and edges.

[0075] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating the working principle of the attention mechanism in the judgment model provided in this application. Figure 4 As shown, the Graph Attention Model (GAT) includes two trainable parameter matrices W1 and W2, a concatenation module (Concat), and a non-linear activation function module (ReLU). After the knowledge graph triples are input into the GAT model, they first pass through the trainable parameter matrices W1 and W2 to obtain the absolute weight values ​​corresponding to each triple, i.e. Figure 4 The beta value is then used; the absolute weight values ​​corresponding to each triple are then input into the softmax module to obtain the relative weight values ​​corresponding to each triple, i.e. Figure 4 The alpha value in the graph is used. Each triplet is associated with its corresponding absolute and relative weight values ​​and stored in list Ai. After K iterations (K is the number of graph updates, generally no more than 3), the graph feature vector representing the current litigation subject point is generated by weighted summation of the triplet vectors (Formula 1).

[0076] h k =∑ N α ，j，p c i，j，p

[0077] Step S14: The first knowledge graph and the second knowledge graph are paired and combined using a double loop method. The judgment model is trained based on the feature vectors between the knowledge graphs to obtain a judgment model for comparing the similarity of any litigation subjects and aligning the judgment subjects with the same subject.

[0078] Specifically, the terminal device employs a dual-loop pairing method to generate knowledge graph pairings. This involves traversing the knowledge graph dataset of each litigation subject within the model training set. On one hand, within each set, training involves traversing the set through pairwise combinations to form multiple knowledge graph pairings for the same litigation subject. On the other hand, cross-set traversal sampling and pairwise pairing are performed between knowledge graph datasets of different subjects to form multiple knowledge graph pairings for different litigation subjects. Since the traversal process knows whether the currently paired subjects are the same subject, a label y can be automatically set for the current pairing. Pairings for the same subject are labeled with 1, while pairings for different subjects are labeled with -1.

[0079] The terminal device inputs the paired feature vectors and label information of the litigation subject knowledge graph into the judgment model, and uses the SoftMarginLoss loss function to control the gradient descent and update of the model parameters to achieve the training of the judgment model.

[0080] Specifically, the terminal device traverses the knowledge graph of litigation subjects for pairing, and generates graph feature vectors for each pair of corresponding two knowledge graphs according to step S13. The inner product of the two feature vectors is used to generate a real number s representing the similarity between the vectors, which represents the probability that the currently paired subjects are the same subject.

[0081] The terminal device uses the SoftMarginLoss loss function, taking input s and the currently paired correct label information y, and calculates the difference between the model prediction result and the true label as the model loss value.

[0082] Since the model loss value reflects the deviation of the current graph attention-based judgment model in recognizing feature vectors of the same and different subject knowledge graphs, updating the judgment model parameters based on the loss value using backpropagation and stochastic gradient descent (SGD) can converge the gap between the judgment model's predicted results and the true labels.

[0083] During the model training process, the terminal device divides the matching data of identical and different subject knowledge graphs into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). The training, validation, and test sets are then mixed to ensure a consistent ratio of identical and different subject knowledge graph data across them. First, the training set data is used to update the graph attention model parameters. Then, the validation set is used to adjust and optimize the model's hyperparameters (e.g., gradient descent learning rate, GAT iteration count K, etc.). Finally, the test set is used to evaluate the performance of the prediction model. When the prediction performance of the model on the test set reaches a threshold, such as an accuracy rate greater than 95% for determining whether two litigants are the same subject, it indicates that the model parameters have been iteratively updated, and the target judgment model is obtained.

[0084] After training to obtain the target judgment model, the trained judgment model can be input into the paired litigation subject knowledge graph to align and fuse litigation subjects judged as the same subject.

[0085] Specifically, for newly input legal litigation subjects, subject knowledge triples are automatically extracted from different data sources using step S11 to construct a knowledge graph describing the subject. The knowledge graph is then paired using the pairing method described in step S13 and input into a graph attention judgment model. The judgment model inputs the current litigation subject similarity s. Based on the specific legal litigation needs, a feature vector similarity threshold s' is determined, and paired subjects with similarities greater than s' are considered as the same litigation subject and aligned and fused.

[0086] Please continue reading. Figure 5 , Figure 5 This is a flowchart illustrating an embodiment of the method for aligning and fusing multi-source heterogeneous legal litigation information provided in this application.

[0087] like Figure 5 As shown, the multi-source heterogeneous legal litigation information alignment and fusion method of this embodiment specifically includes the following steps:

[0088] Step S21: Obtain the target litigation subject type.

[0089] Step S22: Obtain the subject association information corresponding to the target litigation subject type.

[0090] Step S23: Based on the target litigation subject type and its subject relationship information, form several knowledge graphs centered on the target litigation subject type from several legal litigation information data sources.

[0091] Step S24: Use the pre-trained judgment model to match several knowledge graphs and obtain the similarity between each pair of knowledge graphs.

[0092] Step S25: Align the litigation subjects of multiple knowledge graphs with similarity exceeding a preset threshold into the same litigation subject, thereby achieving the fusion of multi-source heterogeneous legal information.

[0093] The judgment model in this application embodiment can be determined by... Figure 1 The model was trained using the method shown, and the training process will not be described in detail here.

[0094] The model training method and multi-source heterogeneous legal litigation information alignment and fusion method provided in this application can extract litigation subjects and their relationships from adapted multi-source heterogeneous data through flexible deep learning information extraction; it overcomes the problem that existing information extraction and natural language processing methods in the field of legal litigation can only handle single data sources and are difficult to effectively identify litigation subject information in multi-source heterogeneous data sources; the graph data format can effectively integrate point and edge semantic information and topological features between subjects and attributes, and quickly judge the similarity of litigation subjects based on feature vectors, overcoming the problem that existing legal litigation subject alignment processes rely on manual labor, are inefficient, and cannot support further legal analysis and judgment; through a double loop method, it matches knowledge graph pairs of the same and different subjects, automatically generates a large amount of knowledge graph pairing training data, and automatically sets whether they are the same subject according to the pairing information, providing sufficient data support for the training of graph attention judgment model and reducing the need for manual annotation.

[0095] To implement the model training method and the multi-source heterogeneous legal litigation information alignment and fusion method in the above embodiments, this application also provides another terminal device 300, for details please refer to Figure 6 The terminal device 300 in this application embodiment includes a processor 31, a memory 32, an input / output device 33, and a bus 34.

[0096] The processor 31, memory 32, and input / output device 33 are respectively connected to the bus 34. The memory 32 stores program data, and the processor 31 is used to execute the program data to implement the model training method and the multi-source heterogeneous legal litigation information alignment and fusion method described in the above embodiments.

[0097] In this embodiment, processor 31 can also be referred to as a CPU (Central Processing Unit). Processor 31 may be an integrated voltage control system chip with signal processing capabilities. Processor 31 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or processor 31 can be any conventional processor.

[0098] This application also provides a computer storage medium; please refer to the following: Figure 7 , Figure 7This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 400 stores program data 41, which, when executed by a processor, is used to implement the model training method and the multi-source heterogeneous legal litigation information alignment and fusion method of the above embodiments.

[0099] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A model training method, characterized in that, The model training method includes: Extract the litigation subjects and their relationships from several heterogeneous multi-source data sources, and construct litigation subject knowledge triples according to the litigation subjects and their relationships; A first knowledge graph is formed by using litigation subject knowledge triples corresponding to the same litigation subject, and a second knowledge graph is formed by using litigation subject knowledge triples corresponding to different litigation subjects. A litigation subject training dataset is then formed based on the first knowledge graph and the second knowledge graph. The knowledge graph within the training dataset of the litigation subjects is vectorized. A dual-loop approach is used to pair and combine the first and second knowledge graphs. A judgment model is trained based on the feature vectors between the knowledge graphs to obtain a judgment model for comparing the similarity of any litigation subjects and aligning the judgments of litigation subjects with the same subject. The method of pairing and combining the first and second knowledge graphs using a dual-loop approach, and training a judgment model based on the feature vectors between the knowledge graphs, includes: Iterate through the knowledge graph of each litigation subject in the training dataset of the litigation subjects. Based on the current knowledge graph, it is paired with other graphs of the same subject and with graphs in knowledge graph datasets of different litigation subjects, and labels are automatically set for graphs of the same subject and graphs of different subjects; In each pair of matching knowledge graphs, the parameters of the judgment model are updated according to the difference between the predicted label and the actual label, thus completing the model training. Specifically, for each pair of corresponding knowledge graphs, a graph feature vector is generated, and the inner product of the two generated graph feature vectors is performed to generate the predicted label representing the vector similarity.

2. The model training method according to claim 1, characterized in that, The extraction of litigation subjects and their relationships from several heterogeneous data sources includes: The litigation subjects and their relationships are extracted from several multi-source heterogeneous data sources using a flexible deep learning information extraction method. The multi-source heterogeneous data source includes structured and unstructured data stored in different data sources and provided through different data interfaces or access methods.

3. The model training method according to claim 2, characterized in that, The flexible deep learning information extraction method includes: constructing a flexible information extraction tool that adapts to the characteristics of the data based on the type of the multi-source heterogeneous data source.

4. The model training method according to claim 1, characterized in that, The construction of the litigation subject knowledge triplet according to the litigation subjects and their relationships includes: The litigation subject is taken as a point in the litigation subject knowledge triple; The relationships between the litigants are used as edges of the litigant knowledge triple; Using two points and the edges connecting the points, construct a triplet of litigation subject knowledge.

5. The model training method according to claim 1, characterized in that, The vectorization representation of the knowledge graph within the training dataset of the litigation subjects includes: Obtain a pre-trained word vector dictionary from an open-source legal litigation text corpus; Based on the semantic names of points and edges in the training dataset of the litigation subjects, the vector dictionary is searched to obtain the semantic vectors of points and edges, thereby realizing the vectorized representation of the knowledge graph. An attention model is used to integrate graph point, edge semantics, and associated topological structure feature information on the litigation subject point to generate a knowledge graph feature vector of the litigation subject.

6. The model training method according to claim 1, characterized in that, After training a judgment model based on feature vectors between knowledge graphs to obtain a judgment model for comparing the similarity of arbitrary litigation subjects and aligning litigation subjects with the same subject, the model training method includes: Obtain any pair of knowledge graphs, wherein the knowledge graphs include a pair of litigation subjects and their corresponding relationships; The judgment model is used to generate knowledge graph vectors of the currently paired litigation subjects and calculate similarity. If a threshold is used to determine whether the paired entities are the same litigant, they will be considered as aligned and merged as if they were the same entity.

7. A method for aligning and fusing multi-source heterogeneous legal litigation information, characterized in that, The information alignment and fusion method includes: Obtain the type of the target litigant; Obtain the subject association information corresponding to the target litigation subject type; Based on the target litigation subject type and its subject relationship information, several knowledge graphs are formed from several legal litigation information data sources, with the target litigation subject type as the central point; The pre-trained judgment model is used to match the knowledge graphs to obtain the pairwise similarity between the knowledge graphs. Aligning the litigants of multiple knowledge graphs with similarity exceeding a preset threshold into the same litigant, thereby achieving the fusion of multi-source heterogeneous legal information; The judgment model is obtained by training the model training method according to any one of claims 1-6.

8. A terminal device, characterized in that, The terminal device includes a processor and a memory connected to the processor, wherein... The memory stores program instructions; The processor is used to execute program instructions stored in the memory to implement the model training method as described in any one of claims 1 to 6 and / or the multi-source heterogeneous legal litigation information alignment and fusion method as described in claim 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores program instructions, which, when executed, implement the model training method as described in any one of claims 1 to 6 and / or the multi-source heterogeneous legal litigation information alignment and fusion method as described in claim 7.

Images