Litigation material processing method, device, equipment, medium and product

Abstract

The invention discloses a litigation material processing method, device and equipment, a medium and a product, and relates to the technical field of data processing. The litigation material processing method comprises the following steps: splitting a classification task into a plurality of subtasks in response to a classification instruction for litigation materials; executing the first subtask to judge whether the name of the litigation material meets a preset naming condition, and if so, classifying according to the name; otherwise, executing a second sub-task to judge whether the litigation materials are of the first data type, and if yes, extracting features from the litigation materials through a classification model and classifying the features; otherwise, executing a third sub-task to judge whether the litigation material is of a second data type, and if the litigation material is of the second data type, identifying the content of the litigation material and classifying the litigation material; and calling a corresponding information extraction agent to extract key information according to the litigation material type to which the litigation material belongs. By adopting the litigation material processing method provided by the invention, the time consumption and manpower in the litigation material sorting and extracting process can be reduced.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, equipment, medium and product for processing litigation materials. Background Technology

[0002] For materials that are diverse in type and have confusing names, the process of organizing and extracting them often relies heavily on manual labor, resulting in low efficiency. For example, in the legal industry, lawyers need to classify and organize complex litigation materials and extract information to assist in the drafting of legal documents such as pleadings. The entire process is time-consuming, and for cases involving a large number of cases or a large amount of materials, the organization and extraction process is both time-consuming and labor-intensive. Summary of the Invention

[0003] This application provides a method, apparatus, equipment, medium, and product for processing litigation materials to address or at least partially address the defects or deficiencies in related technologies.

[0004] Firstly, this application provides a method for processing litigation materials, the method comprising: In response to the classification instruction for the target litigation materials, the classification task for the target litigation materials is broken down into multiple sub-tasks, wherein the multiple sub-tasks include a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type. The first subtask is executed to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material is classified according to its name. If the name of the target litigation material does not meet the preset naming conditions, the second subtask is executed to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the target litigation material is subjected to feature extraction and classification based on the extracted features by a preset classification model for the first data type. If the data type of the target litigation material is not the first data type, the third subtask is executed to determine whether the data type of the target litigation material is the second data type. If the data type of the target litigation material is the second data type, the content of the target litigation material is identified, and the target litigation material is classified according to the identified content. Based on the type of litigation material to which the target litigation material belongs, an information extraction agent preset for the litigation material type is invoked to extract key information from the target litigation material.

[0005] Optionally, in some embodiments of this application, the first data type is an image data type, and the classification model includes an input layer, an image feature extraction layer, and an image classification task layer; the step of extracting features from the target litigation materials using a pre-set classification model for the first data type and classifying the target litigation materials based on the extracted features includes: The target litigation material is received through the input layer, and the target litigation material is preprocessed for size and pixel normalization. Image features are extracted from the target litigation materials through the image feature extraction layer; The image classification task layer classifies the target litigation materials based on the extracted image features.

[0006] Optionally, in some embodiments of this application, the image feature extraction layer includes an initial convolutional layer, multiple cascaded residual block groups, and a global pooling layer, wherein the residual blocks in the residual block groups form residual maps through skip connections; the step of extracting image features from the target litigation material through the image feature extraction layer includes: The initial convolutional layer receives the output of the input layer and extracts the edge color features of the target litigation material. The output of the initial convolutional layer is received through the multiple concatenated residual block groups, and the local texture features, specific scale features, global layout features, and semantic features of the target litigation material are extracted respectively. The global pooling layer receives the outputs of the multiple concatenated residual block groups and performs pooling processing on the outputs of the multiple concatenated residual block groups to obtain the global pooling features of the target litigation material.

[0007] Optionally, in some embodiments of this application, the image classification task layer includes a classification head or a clustering head; the classification of the target litigation materials based on the extracted image features by the image classification task layer includes: The classification head receives the output of the global pooling layer and maps the global pooling features to the litigation material type to obtain the litigation material category to which the target litigation material belongs. or, The clustering head receives the output of the global pooling layer and applies an unsupervised clustering algorithm to the global pooling features to obtain the litigation material category to which the target litigation material belongs.

[0008] Optionally, in some embodiments of this application, the classification model is trained in the following ways: The untrained image feature extraction layer is pre-trained using a pre-defined general image dataset; Collect image-based litigation material datasets; Data augmentation is performed on the aforementioned image litigation material dataset; The input layer is connected to the pre-trained image feature extraction layer, and all the pre-trained weights of the image feature extraction layer are loaded. Freeze the pre-trained weights of the residual block groups except the last residual block group in the concatenated residual block groups, as well as the pre-trained weights of the global pooling layer; Based on the aforementioned image litigation material dataset, the image feature extraction layer is fine-tuned; The untrained image classification task layer is connected to the fine-tuned image feature extraction layer to obtain the classification model to be trained. Freeze all the fine-tuned weights of the image feature extraction layer, or freeze all the fine-tuned weights of the image feature extraction layer except for the fine-tuned weights of the last residual block group; The classification model is trained based on the aforementioned image litigation material dataset.

[0009] Optionally, in some embodiments of this application, determining whether the name of the target litigation material meets the preset naming conditions includes: determining whether the name of the target litigation material conforms to the naming rules corresponding to the preset litigation material type; The step of determining whether the data type of the target litigation material is the first data type includes: determining whether the data type of the target litigation material is an image data type; Determining whether the data type of the target litigation material is the second data type includes: determining whether the data type of the target litigation material is a text data type.

[0010] Secondly, this application provides a litigation material processing device, the litigation material processing device comprising: The task splitting module is configured to respond to the classification instructions for the target litigation materials and split the classification task for the target litigation materials into multiple sub-tasks. The multiple sub-tasks include a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type. The first task execution module is configured to execute the first sub-task to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material is classified according to the name of the target litigation material. The second task execution module is configured to execute the second sub-task if the name of the target litigation material does not meet the preset naming conditions, in order to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the module performs feature extraction on the target litigation material through a preset classification model for the first data type and classifies the target litigation material based on the extracted features. The third task execution module is configured to execute the third sub-task if the data type of the target litigation material is not the first data type, in order to determine whether the data type of the target litigation material is the second data type; if the data type of the target litigation material is the second data type, to identify the content of the target litigation material and to classify the target litigation material according to the identified content. The information extraction module is configured to, based on the type of litigation material to which the target litigation material belongs, invoke a preset information extraction agent for that litigation material type to extract key information from the target litigation material.

[0011] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the litigation material processing method described in the first aspect above.

[0012] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the litigation material processing method described in the first aspect.

[0013] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the litigation material processing method described in the first aspect.

[0014] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, apparatus, device, medium, and product for processing litigation materials. By breaking down the task of classifying litigation materials into multiple sub-tasks, including a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type and feature extraction, and a third sub-task of classifying materials based on a second data type and content recognition, the three sub-tasks can classify litigation materials that meet preset naming conditions, litigation materials that do not meet preset naming conditions but are of the first data type, and litigation materials that do not meet preset naming conditions but are of the second data type, respectively. Then, according to the type of litigation material, the corresponding information extraction agent is invoked to extract key information from the litigation materials, thereby realizing the automated classification and organization of litigation materials and the extraction of key information, reducing the time and manpower required for the process of organizing and extracting litigation materials. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a litigation material processing method provided in an embodiment of this application; Figure 2 A schematic diagram of the functional modules of a litigation material processing device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] In one exemplary embodiment, such as Figure 1As shown, a method for processing litigation materials is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. The terminal can be, but is not limited to, various desktop computers, laptops, smartphones, and tablets. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0020] In the embodiments of this application, such as Figure 1 As shown, the method for processing litigation materials includes the following steps 101 to 105. Wherein: Step 101: In response to the classification instruction for the target litigation materials, the classification task for the target litigation materials is broken down into multiple sub-tasks. The multiple sub-tasks include a first sub-task of classifying materials based on the material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type.

[0021] In litigation, the parties involved include the plaintiff and the defendant. Specific materials related to the parties may include identification documents, credit card application materials, etc. Materials related to the claims may include a calculation of principal and interest, etc. Materials outlining the facts and reasons may include collection records, credit card statements, debt transfer agreements, etc.

[0022] Optionally, the types of litigation materials may include party information materials, litigation request information materials, factual and reason materials, etc. Alternatively, the types of litigation materials may include sub-types such as ID card information, credit card application information, principal and interest calculation list collection records, credit card bills, debt transfer agreements, etc. This application embodiment does not specifically limit these types.

[0023] The target litigation material is litigation material of an unknown type. It can be one of one or more litigation materials selected by the user for categorization. In response to a categorization instruction for the target litigation material, the categorization task can be broken down into multiple sub-tasks, which are then executed in stages and according to different scenarios. These sub-tasks can categorize the target litigation material based on different material elements. Specifically, the first sub-task categorizes materials based on their name, the second sub-task categorizes materials based on a first data type, and the third sub-task categorizes materials based on a second data type. These sub-tasks can automate the categorization and organization of litigation materials with diverse types and inconsistent naming.

[0024] Step 102: Execute the first subtask to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, classify the target litigation material according to its name.

[0025] The first subtask can determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material can be classified according to its name.

[0026] In one optional implementation, determining whether the name of the target litigation material meets preset naming conditions includes: determining whether the name of the target litigation material conforms to the naming rules corresponding to the preset litigation material type. Optionally, the preset litigation material type may include litigation material types such as party information materials, litigation claim information materials, and factual and reasoning materials. For example, the naming rules corresponding to party information materials may include: the litigation material name containing the words "ID card" or "credit card application".

[0027] Among them, corresponding naming rules can be pre-configured for each preset litigation material type. The first subtask can determine whether the name of the target litigation material conforms to the naming rules corresponding to one of the preset litigation material types. If the name of the target litigation material conforms to the naming rules corresponding to a certain preset litigation material type, the target litigation material can be classified into that preset litigation material type and stored in the folder corresponding to that preset litigation material type.

[0028] Step 103: If the name of the target litigation material does not meet the preset naming conditions, execute the second subtask to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, perform feature extraction on the target litigation material through the preset classification model for the first data type and classify the target litigation material based on the extracted features.

[0029] If the name of the target litigation material does not meet the preset naming conditions, the second subtask is executed. The second subtask can determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the target litigation material can be classified using the preset classification model for the first data type.

[0030] In one optional implementation, determining whether the data type of the target litigation material is a first data type includes: determining whether the data type of the target litigation material is an image data type. In this implementation, the first data type is an image data type, that is, determining whether the target litigation material is an image material.

[0031] Optionally, it can be determined whether the data type of the target litigation material is an image data type by checking the file extension of the target litigation material. The file extension can indicate the file format. For example, if the file extension of the target litigation material is .png, .jpg, .jpeg, .bmp, .gif, etc., then it can be determined that the target litigation material is an image data type.

[0032] In this implementation, if the name of the target litigation material does not conform to the naming rules corresponding to any preset litigation material type, a second subtask can be executed. The second subtask can determine whether the target litigation material is an image material. If the target litigation material is an image material, image features are extracted from the target litigation material using a preset classification model (i.e., an image classification model), and the target litigation material is classified based on the extracted image features. The target litigation material is then stored in the folder corresponding to its respective litigation material type.

[0033] To facilitate reading and understanding of the scheme, the optional embodiments of the classification model will be described in detail later.

[0034] Step 104: If the data type of the target litigation material is not the first data type, execute the third subtask to determine whether the data type of the target litigation material is the second data type. If the data type of the target litigation material is the second data type, identify the content of the target litigation material and classify the target litigation material according to the identified content.

[0035] If the data type of the target litigation material is not the first data type, then the third subtask is executed. The third subtask determines whether the data type of the target litigation material is the second data type. If it is, the content of the target litigation material can be identified, and then categorized based on the identified content. The second data type is a different data type from the first data type.

[0036] In one optional implementation, determining whether the data type of the target litigation material is a second data type includes: determining whether the data type of the target litigation material is a text data type. In this implementation, the second data type is a text data type, that is, determining whether the target litigation material is a text material.

[0037] Optionally, it can be determined whether the data type of the target litigation material is a text data type by checking the file extension of the target litigation material. The file extension can indicate the file format. For example, if the file extension of the target litigation material is .doc, .docx, .txt, .xlsx, etc., then it can be determined that the target litigation material is a text data type.

[0038] In this implementation, if the target litigation material is not an image material, a third subtask can be performed. The third subtask can determine whether the target litigation material is a text material. If the target litigation material is a text material, the text content of the target litigation material is identified, the target litigation material is classified according to the identified text content, and the target litigation material is stored in the folder corresponding to the litigation material type.

[0039] Step 105: Based on the type of litigation material to which the target litigation material belongs, invoke the information extraction agent preset for the litigation material type to extract key information from the target litigation material.

[0040] Because the use of intelligent agents can be limited by token restrictions and efficiency issues, corresponding information extraction intelligent agents can be pre-set for different types of litigation materials. After determining the type of litigation material to which the target litigation material belongs, the target litigation material can be routed to the information extraction intelligent agent corresponding to that type of litigation material, thereby extracting key information from the target litigation material. For example, if the target litigation material is ID card information, the key information such as the name and ID number can be extracted from the target litigation material through the information extraction intelligent agent corresponding to ID card information.

[0041] Alternatively, the information extraction agent can employ existing AI agents such as Qwen and Coze.

[0042] Optionally, this step can be implemented in the following ways: Based on the type of litigation material to which the target litigation material belongs, generate prompts for extracting key information. Call the information extraction agent corresponding to the litigation material type to which the target litigation material belongs, and input the target litigation material and prompt words into the information extraction agent to extract the key information in the target litigation material.

[0043] For example, if the target litigation material is an ID card, the prompt for extracting key information could be: "Please extract the name and ID number from this ID card." Then, the information extraction agent corresponding to the ID card material type can be invoked. The target litigation material and the aforementioned prompt can be input into the information extraction agent, which can then extract the name and ID number from the target litigation material.

[0044] For example, if the target litigation material is an ID card, the prompt for extracting key information could be: "Please extract the name and ID number from this ID card." Then, the information extraction agent corresponding to the ID card material type can be invoked. The target litigation material and the aforementioned prompt can be input into the information extraction agent, which can then extract the name and ID number from the target litigation material.

[0045] For example, if the target litigation material is a bond transfer agreement, the prompt for extracting key information could be: "Please extract the name of the transferor, the name of the transferee, and the bond amount from this bond transfer agreement." Then, the information extraction agent corresponding to the bond transfer agreement type can be invoked. The target litigation material and the aforementioned prompt can be input into the information extraction agent, which can then extract the transferor's name, the transferee's name, and the bond amount from the target litigation material.

[0046] Optionally, for litigation materials whose names do not meet the preset naming conditions, if the litigation material is neither a first data type (e.g., not an image material) nor a second data type (e.g., not a text material), it can be regarded as an invalid document and its litigation material category will not be determined. Alternatively, it can enter the manual judgment process to manually determine its litigation material category.

[0047] The litigation material processing method provided in this application decomposes the task of classifying litigation materials into multiple sub-tasks, including a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type and feature extraction, and a third sub-task of classifying materials based on a second data type and content recognition. The three sub-tasks can classify litigation materials that meet preset naming conditions, litigation materials that do not meet preset naming conditions but are of the first data type, and litigation materials that do not meet preset naming conditions but are of the second data type, respectively. Then, according to the type of litigation material, the corresponding information extraction agent is invoked to extract key information from the litigation materials, thereby realizing the automated classification and organization of litigation materials and the extraction of key information, reducing the time and manpower required for the organization and extraction process.

[0048] The following describes a classification model in one alternative embodiment.

[0049] The classification model may include an input layer, an image feature extraction layer, and an image classification task layer. Optionally, the steps of extracting features from the target litigation materials using a pre-defined classification model for the first data type and classifying the target litigation materials based on the extracted features can be implemented through the following steps: Step 1031: Receive the target litigation materials through the input layer, and perform size preprocessing and pixel normalization processing on the target litigation materials.

[0050] The input layer of the classification model receives the input image, i.e., the target litigation material. This input layer performs size preprocessing on the target litigation material, converting input images of different sizes to a uniform size (e.g., a 224×224×3 RGB image) through operations such as scaling up, downscaling, etc. Then, the input layer performs pixel normalization on the preprocessed image, normalizing the pixel values ​​to a specific range, such as [0,1] or [-1,1]. Finally, the input layer inputs the pixel-normalized image into the image feature extraction layer.

[0051] Step 1032: Extract image features from the target litigation materials using the image feature extraction layer.

[0052] For image-based litigation materials (such as scanned copies of newspaper announcements, photos of handwritten application materials, etc.), their visual features include visual layout (such as the position of newspaper headlines, table structure, etc.), texture (such as printed font, handwritten font), and key elements (such as the image area containing the words "debt transfer," etc.). These visual features can be extracted using an image feature extraction layer.

[0053] In one optional implementation, the image feature extraction layer may include an initial convolutional layer, multiple concatenated residual block groups, and a global pooling layer. A residual block group may include multiple residual blocks, which are the basic units for implementing residual operations. The residual blocks within a residual block group can form a residual map through skip connections (also known as shortcut connections). A skip connection in a residual block is a connection method that quickly transmits the input signal to the output. Its design involves processing the input of the residual block through multiple convolutional layers in the main branch, while another shortcut branch quickly transmits the input of the residual block to the output. The outputs of the two branches are then added, and the sum is activated by the ReLU function to ensure the continuity of feature transmission, thus forming a residual map. In this architecture, when the output of the main branch has the same input dimension as the residual block, the outputs of the two branches can be directly added together (i.e., identity mapping). When the output of the main branch has a different input dimension than the residual block, the output of the shortcut branch can be adjusted by convolutional operations to change the number of channels or spatial size before being added to the output of the main branch. The advantage of this structure is that if the main branch cannot learn a better representation, the original input information can be directly preserved, making deep networks easier to optimize.

[0054] In an optional implementation, the step of extracting image features from the target litigation material through an image feature extraction layer can be achieved in the following ways: The initial convolutional layer receives the output of the input layer and extracts the edge color features of the target litigation material. The output of the initial convolutional layer is received by multiple groups of residual blocks connected in series, and local texture features, specific scale features, global layout features and semantic features of the target litigation material are extracted respectively. The global pooling layer receives the outputs of multiple residual block groups in series and performs pooling processing on the outputs of these groups to obtain the global pooling features of the target litigation material.

[0055] The above implementation will be described in detail below with reference to an image feature extraction layer provided by an optional embodiment.

[0056] I. Initial Convolutional Layer: The initial convolutional layer is used to perform preliminary feature extraction on the received image. It can initially extract low-level features such as edges and colors of the image and reduce the image size. The output size of the initial convolutional layer can be 56×56×64.

[0057] The initial convolutional layer includes: 1. Convolutional layer: 7×7 convolutional kernel, 64 filters, stride of 2; 2. Batch Normalization Layer (BatchNorm): Used to improve training stability; 3. ReLU activation layer: used to introduce nonlinearity; 4. Max pooling layer: 3×3 convolution kernel with stride of 2, used to further compress the feature map size.

[0058] Convolutional layers can quickly capture the overall edges and color distribution of an image (such as the black and white printing texture of a newspaper or the background color of handwritten materials). Setting the stride to 2 allows key information to be preserved while reducing the size. Max pooling layers can further reduce the computational cost and enhance local translation invariance (such as slight shifts during material scanning not affecting feature extraction).

[0059] II. Four residual blocks connected in series: 1. The first residual block group: consists of 3 residual blocks connected in series, each with 64 filters. The first residual block has a stride of 2, and the other residual blocks have a stride of 1. The first residual block group is used to extract local texture features (such as differences in the strokes of printed / handwritten characters). The output size of the first residual block group can be 56×56×256. 2. The second residual block group: consists of four residual blocks in series, each with 128 filters. The first residual block has a step size of 2, and the other residual blocks have a step size of 1. The second residual block group is used to extract features at a specific scale (such as table lines and title borders). The output size of the second residual block group can be 28×28×512. 3. The third residual block group: consists of 6 residual blocks in series, each with 256 filters. The first residual block has a step size of 2, and the other residual blocks have a step size of 1. The third residual block group is used to extract global layout features (such as newspaper headline position and material partition structure). The output size of the third residual block group can be 14×14×1024. 4. The fourth residual block group: consists of three cascaded residual blocks, each with 512 filters. The first residual block has a step size of 2, while the other residual blocks have a step size of 1. The fourth residual block group is used to extract high-level semantic features (such as combined features of key regions like "debt transfer"). The output size of the fourth residual block group can be 7×7×2048.

[0060] Each residual block consists of a "main branch" and a "shortcut branch" (if the output dimension of the main branch changes, a 1×1 convolutional kernel can be used to match the output dimension of the main branch; otherwise, a direct identity mapping is used). Skip connections are used to solve the gradient vanishing problem in deep networks. The main branch contains three convolutional layers: the first convolutional layer has a 1×1 kernel (for dimensionality reduction), the second convolutional layer has a 3×3 kernel (for feature extraction), and the third convolutional layer has a 1×1 kernel (for dimensionality increase).

[0061] III. Global Pooling Layer: It contains two convolutional layers, with the first convolutional layer having a kernel size of 7×7 and the second convolutional layer having a kernel size of 1×1.

[0062] Global pooling layers can replace the flattening operation of traditional fully connected layers to compress the dimension of feature maps. For example, a 7×7×2048 feature map can be compressed into a 2048-dimensional vector. This achieves feature dimension compression while preserving feature semantics and avoiding overfitting. It is especially suitable for scenarios in litigation materials where the layout is diverse but the core semantics are consistent, such as different newspapers having different announcement layouts but the same semantic meaning of "debt transfer".

[0063] Step 1033: The target litigation materials are classified based on the extracted image features by the image classification task layer.

[0064] In one optional implementation, the image classification task layer may include a classification head or a clustering head for implementing the downstream task of image feature extraction, i.e., the image classification task. Optionally, the step of classifying the target litigation materials based on the extracted image features by the image classification task layer can be implemented in the following ways: The classification head receives the output of the global pooling layer and maps the global pooling features to the litigation material type to obtain the litigation material category to which the target litigation material belongs. or, The clustering head receives the output of the global pooling layer and applies an unsupervised clustering algorithm to the global pooling features to obtain the category of litigation materials to which the target litigation material belongs.

[0065] The classification head can perform classification tasks, where the categories are known; the clustering head can perform clustering tasks, where the categories are obtained based on the similarity between data.

[0066] The classification head receives the output of the global pooling layer, and then maps the global pooling features to litigation material types. That is, it maps the global pooling features to a certain preset litigation material type among multiple preset litigation material types. The preset litigation material type mapped by the global pooling features is the litigation material category to which the target litigation material belongs.

[0067] Optionally, in one implementation, the classification head may include two fully connected layers. The first fully connected layer can compress high-dimensional features (i.e., globally pooled features), for example, compressing the globally pooled features into a 512-dimensional feature vector. The activation function can be the ReLU function, which introduces non-linearity. The second fully connected layer can use a Softmax classifier to map the features output by the first fully connected layer to a probability distribution of N preset litigation material types, where the preset litigation material type with the highest probability is the litigation material category to which the target litigation material belongs.

[0068] The clustering head receives the output of the global pooling layer and can then apply unsupervised clustering algorithms, such as K-Means (a clustering algorithm based on K-means) and DBSCAN (a clustering algorithm based on density), to the global pooling features, thereby automatically classifying the target litigation materials into the litigation material category to which similar litigation materials belong.

[0069] In one alternative implementation, the classification model is trained in the following manner: Phase 1: Pre-training and fine-tuning of the image feature extraction layer, including: Pre-training: The untrained image feature extraction layer is pre-trained using a pre-defined general image dataset, where the general image dataset refers to a dataset that includes images of various different categories, such as ImageNet. Fine-tuning data preparation: Collect an image litigation material dataset, which includes multiple image litigation material samples of different litigation material types (such as newspaper announcements, handwritten applications, and form-type evidence). Optionally, if the downstream task is a classification task, label each image litigation material sample with the corresponding litigation material type; if the downstream task is a clustering task, simply retain the original data of the image litigation material samples. Data augmentation: Data augmentation is performed on the image litigation material dataset. Optionally, the image litigation material samples can be processed by one or more methods such as random cropping (to adapt to different scanning ratios), rotation (±10° to cope with tilted shooting), brightness / contrast adjustment (to handle lighting differences), and Gaussian blur (to simulate blurred scans), thereby expanding the dataset size and improving the model's generalization ability. Load pre-trained weights: Connect the input layer to the pre-trained image feature extraction layer and load all the pre-trained weights of the image feature extraction layer; Fine-tuning: The pre-trained weights of all residual block groups except the last one, as well as the pre-trained weights of the global pooling layer, are frozen. Then, based on the image litigation material dataset, the image feature extraction layer is fine-tuned. Only the pre-trained weights of the last residual block group and the global pooling layer need to be fine-tuned, thus preserving the general visual feature extraction capability of the image feature extraction layer. The pre-trained weights of other layers remain unchanged, thereby adapting to the high-level layout features of litigation materials. By involving the last residual block group in fine-tuning, it can learn layout features specific to litigation materials, such as "table structure" and "title position".

[0070] The loss function can be optionally configured as follows: if the image litigation material samples are labeled, the cross-entropy loss function can be used (in conjunction with downstream classification tasks); if the image litigation material samples are unlabeled, a self-supervised loss function can be used (in conjunction with downstream clustering tasks), such as a loss function based on contrastive learning, to make the features of similar materials more similar.

[0071] Fine-tuning rounds can be done in 5-10 rounds. When the sample size is small, early stopping can be used to avoid overfitting.

[0072] Phase Two: Downstream task training (classification or clustering tasks), including: Downstream task layer connection: Connect the untrained image classification task layer with the fine-tuned image feature extraction layer to obtain the classification model to be trained. Specifically, an image classification task layer (classification head or clustering head) can be added after the global pooling layer in the image feature extraction layer. Weight Freeze: Freeze all fine-tuned weights of the image feature extraction layer, or freeze all fine-tuned weights of the image feature extraction layer except for the fine-tuned weights of the last residual block group. Model training: The classification model was trained based on the image litigation material dataset.

[0073] For classification tasks, optionally, the optimizer can use SGD (Stochastic Gradient Descent), the learning rate can be set to 1e-3, and the loss function can be the cross-entropy loss function. Training is performed based on sample labels until the accuracy on the validation set is stable. For example, the trained classification head can distinguish between "newspaper announcements" (features: title at the top, multi-column layout) and "handwritten applications" (features: handwritten texture, no fixed layout).

[0074] For clustering tasks, optionally, clustering algorithms, such as the K-Means clustering algorithm, can be applied to the image litigation material dataset (K value is set based on prior knowledge, for example, K=4 according to the type of litigation material). The clustering results can be adjusted through manual verification. For example, if "debt transfer announcements" are clustered into one class, the central features of this class can be calculated as an evaluation criterion for subsequently determining whether a litigation material is a "debt transfer announcement".

[0075] Phase 3: Model evaluation and optimization, including: For classification tasks: you can evaluate classification accuracy and / or confusion matrix (the confusion matrix can visually present the correct and incorrect classification of the classification model in each category), and you can focus on easily confused categories, such as "printed forms" and "handwritten forms".

[0076] For clustering tasks: the silhouette score can be evaluated and / or cluster purity can be calculated using a small number of manually labeled samples. The cluster head can be selected based on the number of clusters with the highest silhouette scores. Optionally, the clustering results obtained during model training and inference can be visualized: high-dimensional features can be reduced to 2D (compressed into a two-dimensional space) using PCA (Principal Component Analysis) for visualization, allowing for a visual representation of the distribution of different types of litigation materials in a two-dimensional chart. Points closer together in the chart (each point representing a piece of litigation material) indicate greater feature similarity. Optionally, folders can be created for each cluster category and corresponding images can be copied; sample images for each cluster category can be displayed to facilitate manual category identification. Optionally, if the clustering effect is not ideal, it can be improved by one or more of the following operations: increasing the number of samples (e.g., adding at least 50 images for each type of litigation material), denoising the samples, adjusting the number of clusters, trying other clustering algorithms (e.g., DBSCAN), and for litigation materials containing text (e.g., scanned debt transfer agreements), text features can be extracted using OCR (Optical Character Recognition) technology and fused with the image features extracted by the image feature extraction layer of the classification model to improve clustering accuracy.

[0077] Optional optimization directions for the classification model: If key elements of the image (such as the words "debt transfer") are not accurately identified, an attention mechanism layer (such as CBAM, Convolutional Block AttentionModule) can be added after the image feature extraction layer to strengthen the feature weights of the text region.

[0078] In summary, this classification model extracts image features in stages through multi-layer convolution and residual connections, gradually abstracting from low-level visual features (texture, edges) to high-level semantic features (layout, key regions), thus adapting to the specific characteristics of litigation materials. The training process involves staged fine-tuning / training, which retains the general capabilities of the pre-trained model while learning the specific features of litigation materials, ultimately achieving efficient classification or clustering tasks for litigation materials.

[0079] Based on the same inventive concept, this application also provides a litigation material processing apparatus for implementing the litigation material processing method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the litigation material processing apparatus provided below can be found in the limitations of the litigation material processing method described above, and will not be repeated here.

[0080] In one exemplary embodiment, such as Figure 2 As shown, a litigation material processing device is provided, which includes: The task splitting module is configured to respond to the classification instructions for the target litigation materials and split the classification task for the target litigation materials into multiple sub-tasks. The multiple sub-tasks include a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type. The first task execution module is configured to execute the first sub-task to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material is classified according to the name of the target litigation material. The second task execution module is configured to execute the second sub-task if the name of the target litigation material does not meet the preset naming conditions, in order to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the module performs feature extraction on the target litigation material through a preset classification model for the first data type and classifies the target litigation material based on the extracted features. The third task execution module is configured to execute the third sub-task if the data type of the target litigation material is not the first data type, in order to determine whether the data type of the target litigation material is the second data type; if the data type of the target litigation material is the second data type, to identify the content of the target litigation material and to classify the target litigation material according to the identified content. The information extraction module is configured to, based on the type of litigation material to which the target litigation material belongs, invoke a preset information extraction agent for that litigation material type to extract key information from the target litigation material.

[0081] Optionally, in some embodiments of this application, the first data type is an image data type, and the classification model includes an input layer, an image feature extraction layer, and an image classification task layer; the step of extracting features from the target litigation materials using a pre-set classification model for the first data type and classifying the target litigation materials based on the extracted features includes: The target litigation material is received through the input layer, and the target litigation material is preprocessed for size and pixel normalization. Image features are extracted from the target litigation materials through the image feature extraction layer; The image classification task layer classifies the target litigation materials based on the extracted image features.

[0082] Optionally, in some embodiments of this application, the image feature extraction layer includes an initial convolutional layer, multiple cascaded residual block groups, and a global pooling layer, wherein the residual blocks in the residual block groups form residual maps through skip connections; the step of extracting image features from the target litigation material through the image feature extraction layer includes: The initial convolutional layer receives the output of the input layer and extracts the edge color features of the target litigation material. The output of the initial convolutional layer is received through the multiple concatenated residual block groups, and the local texture features, specific scale features, global layout features, and semantic features of the target litigation material are extracted respectively. The global pooling layer receives the outputs of the multiple concatenated residual block groups and performs pooling processing on the outputs of the multiple concatenated residual block groups to obtain the global pooling features of the target litigation material.

[0083] Optionally, in some embodiments of this application, the image classification task layer includes a classification head or a clustering head; the classification of the target litigation materials based on the extracted image features by the image classification task layer includes: The classification head receives the output of the global pooling layer and maps the global pooling features to the litigation material type to obtain the litigation material category to which the target litigation material belongs. or, The clustering head receives the output of the global pooling layer and applies an unsupervised clustering algorithm to the global pooling features to obtain the litigation material category to which the target litigation material belongs.

[0084] Optionally, in some embodiments of this application, the classification model is trained in the following ways: The untrained image feature extraction layer is pre-trained using a pre-defined general image dataset; Collect image-based litigation material datasets; Data augmentation is performed on the aforementioned image litigation material dataset; The input layer is connected to the pre-trained image feature extraction layer, and all the pre-trained weights of the image feature extraction layer are loaded. Freeze the pre-trained weights of the residual block groups except the last residual block group in the concatenated residual block groups, as well as the pre-trained weights of the global pooling layer; Based on the aforementioned image litigation material dataset, the image feature extraction layer is fine-tuned; The untrained image classification task layer is connected to the fine-tuned image feature extraction layer to obtain the classification model to be trained. Freeze all the fine-tuned weights of the image feature extraction layer, or freeze all the fine-tuned weights of the image feature extraction layer except for the fine-tuned weights of the last residual block group; The classification model is trained based on the aforementioned image litigation material dataset.

[0085] Optionally, in some embodiments of this application, determining whether the name of the target litigation material meets the preset naming conditions includes: determining whether the name of the target litigation material conforms to the naming rules corresponding to the preset litigation material type; The step of determining whether the data type of the target litigation material is the first data type includes: determining whether the data type of the target litigation material is an image data type; Determining whether the data type of the target litigation material is the second data type includes: determining whether the data type of the target litigation material is a text data type.

[0086] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3 As shown, the computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a litigation material processing method.

[0087] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0088] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0089] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0090] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0091] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0093] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for processing litigation materials, characterized in that, The methods for processing litigation materials include: In response to the classification instruction for the target litigation materials, the classification task for the target litigation materials is broken down into multiple sub-tasks, wherein the multiple sub-tasks include a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type. The first subtask is executed to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material is classified according to its name. If the name of the target litigation material does not meet the preset naming conditions, the second subtask is executed to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the target litigation material is subjected to feature extraction and classification based on the extracted features by a preset classification model for the first data type. If the data type of the target litigation material is not the first data type, the third subtask is executed to determine whether the data type of the target litigation material is the second data type. If the data type of the target litigation material is the second data type, the content of the target litigation material is identified, and the target litigation material is classified according to the identified content. Based on the type of litigation material to which the target litigation material belongs, an information extraction agent preset for the litigation material type is invoked to extract key information from the target litigation material.

2. The method for processing litigation materials according to claim 1, characterized in that, The first data type is an image data type, and the classification model includes an input layer, an image feature extraction layer, and an image classification task layer; the step of extracting features from the target litigation materials using a pre-set classification model for the first data type and classifying the target litigation materials based on the extracted features includes: The target litigation material is received through the input layer, and the target litigation material is preprocessed for size and pixel normalization. Image features are extracted from the target litigation materials through the image feature extraction layer; The image classification task layer categorizes the target litigation materials based on the extracted image features.

3. The method for processing litigation materials according to claim 2, characterized in that, The image feature extraction layer includes an initial convolutional layer, multiple concatenated residual block groups, and a global pooling layer, wherein the residual blocks in the residual block groups form residual maps through skip connections; the image feature extraction of the target litigation material through the image feature extraction layer includes: The initial convolutional layer receives the output of the input layer and extracts the edge color features of the target litigation material. The output of the initial convolutional layer is received through the multiple concatenated residual block groups, and the local texture features, specific scale features, global layout features, and semantic features of the target litigation material are extracted respectively. The global pooling layer receives the outputs of the multiple concatenated residual block groups and performs pooling processing on the outputs of the multiple concatenated residual block groups to obtain the global pooling features of the target litigation material.

4. The litigation material processing method according to claim 2, characterized in that, The image classification task layer includes a classification head or a clustering head; the process of classifying the target litigation materials based on extracted image features by the image classification task layer includes: The classification head receives the output of the global pooling layer and maps the global pooling features to the litigation material type to obtain the litigation material category to which the target litigation material belongs. or, The clustering head receives the output of the global pooling layer and applies an unsupervised clustering algorithm to the global pooling features to obtain the litigation material category to which the target litigation material belongs.

5. The litigation material processing method according to claim 3, characterized in that, The classification model is trained in the following ways: The untrained image feature extraction layer is pre-trained using a pre-defined general image dataset; Collect image-based litigation material datasets; Data augmentation is performed on the aforementioned image litigation material dataset; The input layer is connected to the pre-trained image feature extraction layer, and all the pre-trained weights of the image feature extraction layer are loaded. Freeze the pre-trained weights of the residual block groups except the last residual block group in the concatenated residual block groups, as well as the pre-trained weights of the global pooling layer; Based on the aforementioned image litigation material dataset, the image feature extraction layer is fine-tuned; The untrained image classification task layer is connected to the fine-tuned image feature extraction layer to obtain the classification model to be trained. Freeze all the fine-tuned weights of the image feature extraction layer, or freeze all the fine-tuned weights of the image feature extraction layer except for the fine-tuned weights of the last residual block group; The classification model is trained based on the aforementioned image litigation material dataset.

6. The method for processing litigation materials according to claim 1, characterized in that, The step of determining whether the name of the target litigation material meets the preset naming conditions includes: determining whether the name of the target litigation material conforms to the preset naming rules corresponding to the litigation material type; The step of determining whether the data type of the target litigation material is the first data type includes: determining whether the data type of the target litigation material is an image data type; Determining whether the data type of the target litigation material is the second data type includes: determining whether the data type of the target litigation material is a text data type.

7. A litigation material processing device, characterized in that, The litigation material processing device includes: The task splitting module is configured to respond to the classification instructions for the target litigation materials and split the classification task for the target litigation materials into multiple sub-tasks. The multiple sub-tasks include a first sub-task of classifying materials based on material name, a second sub-task of classifying materials based on a first data type, and a third sub-task of classifying materials based on a second data type. The first task execution module is configured to execute the first sub-task to determine whether the name of the target litigation material meets the preset naming conditions. If the name of the target litigation material meets the preset naming conditions, the target litigation material is classified according to the name of the target litigation material. The second task execution module is configured to execute the second sub-task if the name of the target litigation material does not meet the preset naming conditions, in order to determine whether the data type of the target litigation material is the first data type. If the data type of the target litigation material is the first data type, the module performs feature extraction on the target litigation material through a preset classification model for the first data type and classifies the target litigation material based on the extracted features. The third task execution module is configured to execute the third sub-task if the data type of the target litigation material is not the first data type, in order to determine whether the data type of the target litigation material is the second data type; if the data type of the target litigation material is the second data type, to identify the content of the target litigation material and to classify the target litigation material according to the identified content. The information extraction module is configured to, based on the type of litigation material to which the target litigation material belongs, invoke a preset information extraction agent for that litigation material type to extract key information from the target litigation material.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the litigation material processing method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the litigation material processing method according to any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the litigation material processing method according to any one of claims 1-6.

Images