A method and device for identifying case dispute focus based on multi-feature fusion

Through the multi-feature fusion method, pre-trained language models, syntactic dependency graphs and knowledge graphs are used to generate hyperbolic manifold representations, which solves the problems of low text interpretability and high computing resource consumption in the identification of case dispute focus, and achieves efficient and accurate real-time recognition.

CN116226711BActive Publication Date: 2025-09-09CHONGQING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310166484.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-02-24
Publication Date
2025-09-09
Estimated Expiration
2043-02-24

AI Technical Summary

Technical Problem

Existing technologies for identifying case dispute focus have problems such as low text interpretability, high consumption of computing resources, and inaccurate recognition results, especially when processing long text data, making it difficult to achieve real-time prediction.

Method used

The pre-trained language model encoding is used to generate semantic hyperbolic manifold representation, a syntactic dependency graph is constructed and mapped to a syntactic hyperbolic manifold representation, the knowledge graph is linked to generate a knowledge hyperbolic manifold representation, and prediction is performed by adaptively fusing multiple hyperbolic manifold representations.

Benefits of technology

It improves the model's interpretability and recognition accuracy, reduces computing resource consumption, and enables real-time prediction and efficient identification of dispute focus.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116226711B_ABST
    Figure CN116226711B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and device for identifying case dispute focal points based on multi-feature fusion. The method comprises: first, encoding and mapping a semantic hyperbolic manifold representation using a pre-trained language model; constructing a syntactic dependency graph to obtain a text vector representation and mapping it to a syntactic hyperbolic manifold representation; then, generating a knowledge hyperbolic manifold representation by linking knowledge graphs; and finally, adaptively fusing multiple representations and performing joint reasoning to predict dispute focal points. By designing an adaptive feature fusion model based on hyperbolic space, the present invention integrates text semantic information, syntactic information, and knowledge elements, thereby improving model interpretability, enhancing text features, and reducing the difficulty of distinguishing easily confused categories. The model is trained using negative sampling, reducing computing resources, achieving dispute focal point identification, and improving trial efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a method and device for identifying case dispute focus based on multi-feature fusion. Background Art

[0002] Current research on dispute focus identification based on deep learning primarily utilizes text summarization and multi-label text classification techniques. Text summarization primarily generates summaries by learning high-order features from text. However, in real-world data, it often only produces short summaries. Generating long text summaries also presents challenges with model convergence and interpretability. Multi-label text classification identifies dispute focus by establishing a dispute focus database and classifying pairs of sentences. However, in practical applications, this often results in low recognition accuracy due to imbalanced classification. Graph neural networks are easily applicable to inductive problems and can achieve real-time learning and classification of new nodes. However, dispute focus identification data typically consists of long text. Processing this data as a text graph results in highly sparse text features, which are prone to representation collapse. Transmitting messages through the graph structure typically consumes significant computing resources, making it difficult to achieve real-time prediction.

[0003] To address the low interpretability of text, researchers are using knowledge graphs as a supplement to help summarize the key points of contention between the plaintiff and the defense. However, judicial data itself is limited by the length of text, and simply concatenating features increases the difficulty of processing. Additionally, some researchers are combining dependency parsing techniques with text classification techniques to improve model interpretability. However, text features containing only syntactic information often limit the effectiveness of text classification. Summary of the Invention

[0004] In view of this, it is necessary to design a method and device for identifying the focus of case disputes based on multi-feature fusion. First, the semantic hyperbolic manifold representation is obtained by encoding and mapping the pre-trained language model, and the syntactic dependency graph is constructed to obtain the text vector representation and mapped to the syntactic hyperbolic manifold representation. Then, the knowledge hyperbolic manifold representation is generated by linking the knowledge graph. Finally, multiple hyperbolic manifold representations are adaptively fused to predict the focus of the case statement data to be identified.

[0005] The technical solutions adopted by the present invention to solve the above technical problems include:

[0006] In a first aspect of the present invention, the present invention provides a method for identifying case dispute focus based on multi-feature fusion, the method comprising:

[0007] Pre-process the Chinese data of the case pleading statement data to be identified;

[0008] Use the pre-trained language model to encode the first text vector representation and map it to a semantic hyperbolic manifold representation;

[0009] Construct a syntactic dependency graph to obtain the second text vector representation and map it to a syntactic hyperbolic manifold representation;

[0010] The third text vector representation is obtained by linking the knowledge graph and mapped into a knowledge hyperbolic manifold representation;

[0011] Adaptively fuse multiple hyperbolic manifold representations to predict the controversial focus of the case statement data to be identified.

[0012] In a second aspect of the present invention, the present invention further provides a device for identifying case dispute focus based on multi-feature fusion, which is applied to the method of the first aspect of the present invention, characterized in that the device includes a data acquisition module, a data processing module, a semantic encoding module, a syntactic encoding module, a knowledge graph module, a data mapping module, and an adaptive fusion module;

[0013] The data acquisition module is used to acquire the case pleading statement data to be identified;

[0014] The data processing module is used to pre-process Chinese data to obtain legal text data;

[0015] The semantic encoding module is used to encode the processed text through a pre-trained language model to obtain a first text vector representation;

[0016] The syntactic encoding module is used to generate a second text vector representation by subjecting the first text vector representation to syntactic dependency analysis;

[0017] The knowledge graph module is used to obtain a third text vector representation by linking the knowledge graph;

[0018] The data mapping module is used to map the first text vector representation, the second text vector representation, and the third text vector representation to a hyperbolic space to generate corresponding hyperbolic manifold representations;

[0019] The adaptive fusion module is used to fuse multiple hyperbolic manifold representations of the hyperbolic space and predict the controversial issues involved in the case.

[0020] Beneficial effects of the present invention:

[0021] The present invention transforms the problem of identifying focus of dispute in the judicial field into a multi-label text classification task. In order to solve the problem of differences between the corpus used by general pre-trained language models and legal texts, an adaptive feature fusion model based on hyperbolic space is designed to integrate text semantic information, syntactic information and knowledge elements, thereby improving the interpretability of the model, enhancing text features, and reducing the difficulty of distinguishing easily confused categories. The model is trained by negative sampling, which reduces computing resources, realizes the identification of focus of dispute, enhances the real-time performance and accuracy of model prediction, and has better results. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a schematic diagram of the process of identifying the focus of dispute in a case based on multi-feature fusion according to an embodiment of the present invention;

[0023] Figure 2 Schematic diagram of the architecture of the case dispute focus identification model based on multi-feature fusion described in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] like Figures 1-2 As shown, a method for identifying case dispute focus based on multi-feature fusion includes but is not limited to the following steps:

[0026] 101. Pre-process the Chinese data of the case pleading statement data to be identified;

[0027] Before performing Chinese data preprocessing on the case pleadings data to be identified, it is necessary to obtain the case pleadings data for Chinese data preprocessing. The case pleadings data to be identified can be understood as unfinished case pleadings data or judgment document data. When training the dispute focus identification module, the judgment document data can be used as the training set data. When verifying or testing the dispute focus identification module, the unfinished case pleadings data can be used as the validation set data or the test set data. Data preprocessing includes using the Jieba word segmentation tool to segment legal text, using Stanford's official pre-trained Glove word vectors for knowledge embedding and label embedding, and removing duplicate and excessively short text data. Special symbols and English characters are also removed.

[0028] 102. Use the pre-trained language model to encode and obtain a first text vector representation, and map it into a semantic hyperbolic manifold representation;

[0029] In one embodiment, the pre-trained language model BERT predicts missing words in the text through an automatic random masking mechanism, and uses the Next Sentence Prediction (NSP) task to jointly express the sentence sequence representation of the text. BERT is trained by maximizing the likelihood function of the predicted words, and the BERT model is applied to learn the vector representation of the text, and then adapted to the multi-label text classification task mode through fine-tuning. First, the processed text i is segmented using the Jieba word segmenter, and then the word is first converted into a tag embedding vector, a position embedding vector, and a segmentation embedding vector using the embedding layer of BERT. Then, the vector representation of the word is generated using the encoding layer of BERT, and all the word vector representations are combined into a vector to represent the text. The encoding layer of BERT contains an encoder with 12 layers of Transformer architecture stacked on top of each other. Each Transformer architecture implements a multi-head attention mechanism and contains a feedforward neural network with a normalization layer. High-level Transformers tend to encode text-level feature information. Therefore, this paper uses the output of the penultimate layer of the BERT encoding layer as the first text vector representation, and finally maps the first text vector representation to a semantic hyperbolic manifold representation h i .

[0030] Hyperbolic space is a space with negative constant curvature, meaning that the curvature at any location in the space is negative. Many models can be used to describe hyperbolic space, the most common being the Poincaré disk model. Every hyperbolic space can be described by a corresponding Poincaré disk. Exponential and logarithmic transformations can map the local hyperbolic space at a point to its tangent space. The logarithmic transformation maps the point to the tangent space, and the exponential transformation maps it back to the hyperbolic space.

[0031] Specifically, the exponential mapping function is expressed as:

[0032]

[0033] Among them, t i Represents the first text vector representation of text i encoded by the pre-trained language model BERT, that is, the semantic text feature in Euclidean space, m is represented as a point in hyperbolic space, Expressed as the tangent space at point m, c represents the negative curvature constant Expressed as using a radius of The Poincare sphere model, Represented as the first text vector representation t at point m i Perform exponential mapping, make When m=0:

[0034]

[0035] in, represents the Möbius addition, You can get the characteristics of the node in the hyperbolic space, In order to obtain better representation capability, a linear transformation method parameterized by the weight sharing matrix M is used to convert the hyperbolic features into a higher-order representation. Unlike the matrix-vector multiplication in Euclidean space, Möbius multiplication is required in the hyperbolic space.

[0036]

[0037] Among them, h i represents the semantic hyperbolic manifold feature of node i, and M represents the weight sharing matrix; represents the Möbius multiplication, Represents the first text vector at point 0, t i The features after exponential mapping, represents the radius of the Poincare sphere model. Möbius multiplication satisfies the matrix associative law and if Mq i =0, then h i As the representation of text vector i in hyperbolic space, that is, semantic hyperbolic manifold representation.

[0038] 103. Construct a syntactic dependency graph to obtain a second text vector representation, and map it into a syntactic hyperbolic manifold representation;

[0039] In one embodiment, a Stanford syntactic analyzer is first used to perform syntactic dependency analysis on sentences in a text. The text is converted into a graph using words as nodes and syntactic dependencies as edges. A graph convolutional network is then used to learn the text topology structure and aggregate word representations. The generated text vector representation is then mapped to a hyperbolic space using an exponential mapping function and converted into a high-order syntactic hyperbolic manifold representation.

[0040] Specifically, the syntactic dependency graph G = (V, E) is input into the graph convolutional network to obtain the word aggregation representation w v :

[0041] w v =σ(mlp(H l+1 ))

[0042] H l+1 =AH lW

[0043] Where V represents a word node, E represents an edge generated by syntactic dependency; w v is the final word embedding representation output by the graph convolutional network; σ(.) is the nonlinear transformation function; mlp is the fully connected layer. A is the adjacency matrix of the syntactic graph; H l is the embedding matrix output by the lth layer of the syntactic graph; H 0 is the word feature input to the first layer of the graph convolutional network; W is the trainable weight matrix;

[0044] Use mean pooling to generate the second text vector representation s i :

[0045]

[0046] Where |v| represents the word node representation; G i Represents the set of all word nodes of text i; for the second text vector s i Perform mapping to generate syntactic hyperbolic manifold representation:

[0047]

[0048] 104. Obtain the third text vector representation by linking the knowledge graph and mapping it into a knowledge hyperbolic manifold representation;

[0049] In one embodiment, the words in the text i are first associated with the knowledge graph, that is, the knowledge triples involved in all entities in the text are found from the legal knowledge graph through knowledge query to generate knowledge word embeddings, and the knowledge word embeddings are compressed and transformed using a feedforward neural network (FFN) composed of a double-layer fully connected layer and an activation function, and then converted into a hyperbolic knowledge representation using an exponential mapping.

[0050] Specifically, the knowledge words involved in text i are embedded and transformed using FFN to generate knowledge representation j i , and then mapped to the hyperbolic space to generate the hyperbolic knowledge representation:

[0051] 105. Adaptively integrate multiple hyperbolic manifold representations to predict the controversial focus of the case statement data to be identified;

[0052] Furthermore, in one embodiment, the semantic hyperbolic manifold representation, the syntactic hyperbolic manifold representation, and the knowledge hyperbolic manifold representation are input into the adaptive fusion module for reasoning. Specifically, the three features are fused using a hyperbolic nonlinear transformation:

[0053]

[0054]

[0055] where z i represents the fused text representation; c and c' are the hyperbolic space curvatures of the current layer and the next layer, respectively. Since the two hyperbolic spaces share the tangent space at point 0, the curvatures of the two layers can be smoothed during the nonlinear activation process; φ(·) is the hyperbolic nonlinear transformation; Indicates logarithmic mapping at node 0; Indicates exponential mapping at node 0; W h 、W g 、W j is the learnable parameter matrix, h i is a semantic hyperbolic manifold representation; g i is the syntactic hyperbolic manifold representation; k i It is a hyperbolic manifold representation of knowledge. Each text often contains multiple labels. Multiple label words are embedded and mapped into the hyperbolic space to generate label hyperbolic embedding. Specifically, the fused text representation z i and the lth label embedding y l The probability of the predicted text containing the lth label input into the interaction function sim(.,.) is:

[0056]

[0057] Among them, f D (d)=0.5-arctan(8d-20) / π,f D :[0,∞]→[0,1], and satisfy monotonically decreasing, f D (0) = 1, f D (∞)=0, d represents the fused hyperbolic text representation z i With the hyperbolic label semantic feature y l The geodesic distance of . The value of similarity below the threshold ε is recorded as 0.

[0058] It can be understood that the hyperbolic label semantic feature y l The same mapping method as the fused hyperbolic text representation is used, that is, the corresponding hyperbolic label semantic features can be obtained by mapping the label information vector into the hyperbolic space.

[0059] In the above interactive process, computational complexity increases as the number of labels increases. When the number of labels is too large, calculating the text representation that adapts to the labels requires a huge amount of computing resources. Considering that each text usually has only a small number of labels, negative sampling is used to reduce the number of labels required during training. The loss function is:

[0060]

[0061] Among them, L + represents the set of true labels, L - Represents the set of negative example labels obtained by negative sampling. Output the probability of each label and predict the focus of controversy.

[0062] The Adam optimizer is used to minimize the loss function, early stopping is used to avoid overfitting, and training is iterated until the model converges.

[0063] In some embodiments of the present invention, the present invention provides a device for identifying case dispute focus based on multi-feature fusion, which is applied to the method of the first aspect of the present invention, the device comprising a data acquisition module, a data processing module, a semantic encoding module, a syntactic encoding module, a knowledge graph module, a data mapping module, and an adaptive fusion module;

[0064] The data acquisition module is used to acquire the case pleading statement data to be identified;

[0065] The data processing module is used to pre-process Chinese data to obtain legal text data;

[0066] The semantic encoding module is used to encode the processed text through a pre-trained language model to obtain a first text vector representation;

[0067] The syntactic encoding module is used to generate a second text vector representation by subjecting the first text vector representation to syntactic dependency analysis;

[0068] The knowledge graph module is used to obtain a third text vector representation by linking the knowledge graph;

[0069] The data mapping module is used to map the first text vector representation, the second text vector representation, and the third text vector representation to a hyperbolic space to generate corresponding hyperbolic manifold representations;

[0070] The adaptive fusion module is used to fuse multiple hyperbolic manifold representations of the hyperbolic space and predict the controversial issues involved in the case.

[0071] In some embodiments of the present invention, a terminal for a method for identifying the focus of case disputes based on multi-feature fusion is also provided, comprising a processor and a memory, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, the method described in the above-mentioned method for identifying the focus of case disputes based on multi-feature fusion is implemented.

[0072] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: ROM, RAM, disk or CD, etc.

[0073] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying case dispute focus based on multi-feature fusion, characterized by: The following steps are involved: Pre-process the Chinese data of the case pleading statement data to be identified; The Chinese data preprocessing includes organizing the case pleading and defense statement data to be identified into the format of pleading content and label sets according to a predefined dispute focus label system, deleting duplicate and short data after the organization; deleting garbled characters and English characters in the text, and removing meaningless stop words; counting the sentences with a length greater than 512 characters from the processed text, splitting them into the plaintiff's claim and the defendant's claim, and labeling them with the same label; Use the pre-trained language model to encode the first text vector representation and map it to a semantic hyperbolic manifold representation; Construct a syntactic dependency graph to obtain the second text vector representation and map it to a syntactic hyperbolic manifold representation; The third text vector representation is obtained by linking the knowledge graph and mapped into a knowledge hyperbolic manifold representation; Adaptively fuse multiple hyperbolic manifold representations to predict the controversial focus of the case statement data to be identified; The adaptive fusion of multiple knowledge hyperbolic manifold representations predicts the controversial focus of the case statement data to be identified, including: The semantic hyperbolic manifold representation, syntactic hyperbolic manifold representation, and knowledge hyperbolic manifold representation are input into the adaptive fusion module for reasoning. The three hyperbolic manifold features are fused using hyperbolic nonlinear transformation, and negative sampling is used to reduce the amount of computation. The hyperbolic nonlinear transformation includes: Among them, z i represents the fused text representation; φ is the hyperbolic nonlinear transformation; c and c' are the hyperbolic space curvature of the current layer and the hyperbolic space curvature of the next layer respectively; represents the Möbius addition, Indicates logarithmic mapping at node 0; W h 、W g 、W j is the learnable parameter matrix, h i is a semantic hyperbolic manifold representation; g i is the syntactic hyperbolic manifold representation; k i It is a hyperbolic manifold representation of knowledge.

2. The method for identifying case dispute focus based on multi-feature fusion according to claim 1 is characterized in that: The method of using a pre-trained language model to encode and obtain a first text vector representation includes: using the Jieba word segmenter to segment the processed text, and then using the BERT embedding layer to convert the words into tag embedding vectors, position embedding vectors and segmentation embedding vectors; using the BERT encoding layer to generate vector representations of the words, and combining all the word vector representations into a vector to represent the text; wherein the BERT encoding layer includes an encoder with multiple layers of Transformer architectures stacked together, each Transformer architecture implements a multi-head attention mechanism and includes a feedforward neural network with a normalization layer, and the output of the penultimate layer of the BERT encoding layer is used as the first text vector representation.

3. The method for identifying case dispute focus based on multi-feature fusion according to claim 1 is characterized in that: Constructing a syntactic dependency graph to obtain a second text vector representation includes: Use Stanford's syntactic analyzer to perform syntactic dependency analysis on sentences in the text, convert the text into a graph with words as nodes and syntactic dependencies as edges, and then use a graph convolutional network to learn the text topology structure and aggregate word representations to generate a second text vector representation.

4. The method for identifying case dispute focus based on multi-feature fusion according to claim 1 is characterized in that: Generating a hyperbolic knowledge representation by linking knowledge graphs includes: The words in the text are associated with the knowledge graph. That is, through knowledge query, the knowledge triples involved in all entities in the text are found from the legal knowledge graph to generate knowledge word embeddings. The knowledge word embeddings are compressed and transformed using a feedforward neural network composed of a double-layer fully connected layer and an activation function to generate a third text vector representation.

5. The method for identifying case dispute focus based on multi-feature fusion according to claim 1, 2, 3 or 4, characterized in that: Mapping to a hyperbolic manifold representation includes using an exponential mapping function to map the corresponding text vector to a hyperbolic space, and using a linear transformation method parameterized by a weight sharing matrix to convert the hyperbolic features into a higher-order hyperbolic manifold representation.

6. A device for identifying the focus of dispute in a case based on multi-feature fusion, which is applied to the method for identifying the focus of dispute in a case based on multi-feature fusion as claimed in any one of claims 1 to 5, characterized in that: The device includes a data acquisition module, a data processing module, a semantic encoding module, a syntactic encoding module, a knowledge graph module, a data mapping module, and an adaptive fusion module; The data acquisition module is used to acquire the case pleading statement data to be identified; The data processing module is used to pre-process Chinese data to obtain legal text data; The semantic encoding module is used to encode the processed text through a pre-trained language model to obtain a first text vector representation; The syntactic encoding module is used to generate a second text vector representation by subjecting the first text vector representation to syntactic dependency analysis; The knowledge graph module is used to obtain a third text vector representation by linking the knowledge graph; The data mapping module is used to map the first text vector representation, the second text vector representation, and the third text vector representation to a hyperbolic space to generate corresponding hyperbolic manifold representations; The adaptive fusion module is used to fuse multiple hyperbolic manifold representations of the hyperbolic space and predict the controversial issues involved in the case.