Legal element-based semantic syntactic fusion violation prediction method

Abstract

The embodiment of the invention provides a semantic syntactic fusion violation prediction method based on legal elements, and belongs to the field of intelligent legal analysis, and the method specifically comprises the steps: 1, coding legal constituent elements, and generating differential feature representation; step 2, acquiring case text data, extracting semantic and syntactic information, and constructing semantic-syntactic fusion representation; 3, adopting a multi-hop reasoning technology to construct a cross-sentence dependency relationship network, and capturing a complex logic relationship in the legal text; 4, constructing a cross attention weight based on the sentence-level global relation matrix, dynamically adjusting the embedding weight of the core sentence, and improving the key information characterization capability; and 5, dynamically adjusting a prediction strategy through a reinforcement learning reward mechanism, optimizing sentence selection in combination with a historical selection record and a diversity penalty term, and enhancing an LCE automatic matching capability in combination with a legal knowledge base constraint.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent legal reasoning, and particularly to a semantic-syntactic fusion violation prediction method based on legal elements. Background Art

[0002] Legal clause violation prediction, as the core task of intelligent legal reasoning, aims to automatically identify potential criminal or civil violations through case descriptions. In the civil law system, this technology needs to strictly follow the legal element matching logic chain of "objective facts → subject → subjective mental state → object". However, existing methods have significant limitations: relying on single-feature modeling, it is difficult to handle the complex cross-sentence dependencies of legal texts. Traditional technologies are divided into two categories: although semantic analysis methods can extract text features, they ignore the dependency relationships and logical structures between elements, resulting in insufficient multi-element collaborative judgment; although graph structure methods can construct relationship graphs, the node semantics are mismatched due to artificially defined relationship types, and single-layer graph convolution cannot capture cross-sentence long-distance dependencies, restricting multi-hop reasoning capabilities.

[0003] Currently, CN113487453A strengthens element extraction through hierarchical Bi-LSTM, but lacks syntactic constraint modeling and cannot parse the "motive-behavior" logic chain; CN118484536A introduces a graph convolution network to enhance structural analysis, but is limited by the reasoning defects of artificial relationship definition and single-layer architecture. The two types of methods together reflect the bottleneck of semantic-syntactic feature fragmentation and insufficient cross-sentence logical integration. Therefore, there is an urgent need for a new solution that integrates multi-dimensional features, supports cross-sentence multi-hop reasoning, and embeds legal rule constraints to break through the technical barriers of complex legal text analysis. Summary of the Invention

[0004] In view of the fact that the logical relationship between legal components is the core basis for legal violation prediction, the present invention intends to design a dynamic matching mechanism that can effectively improve the accuracy of legal provision violation prediction by combining multi-hop syntactic enhancement and semantic fusion methods.

[0005] Currently, mainstream methods adopt single-feature modeling: semantic analysis methods construct global features through attention mechanisms, but cannot parse the conditional dependency relationships between legal elements; syntactic analysis methods rely on artificially defined dependency relationships and are difficult to capture element associations across sentences. In addition, existing technologies often ignore the close connection between syntax and semantics, resulting in limited prediction accuracy. To break through the existing bottleneck, the present invention proposes to construct a multi-hop syntactic dependency tree through dynamic data-driven, and combine a semantic perception module and a syntactic perception module to systematically capture cross-sentence dependency relationships and the complex interaction between semantic and syntactic features, thereby effectively improving the modeling ability of complex relationships in legal texts.

[0006] The technical solution provided by the present invention includes the following main steps: Step 1, legal constituent element encoding; Step 2, case document reading; Step 3, semantic and syntactic fusion; Step 4, multi-hop syntactic reasoning; Step 5, automatic legal agent matching and violation prediction.

[0007] Step 1: Through the Legal Constituent Element Encoder (LCE Encoder), generate differentiated feature representations for each legal constituent element, and automatically extract legal elements in the case text. Use the GRU network to perform context encoding on the text sequence of each legal constituent element, and extract basic features through max pooling. Combine the variance-driven dynamic screening mechanism to enhance the quality of feature representation.

[0008] Step 2: Obtain the case text data to be processed. First, use the HanLP tool to perform syntactic dependency analysis on the legal text, extract key syntactic paths such as subject-predicate relationship (nsubj) and verb-object relationship (dobj), and construct a syntactic dependency tree to avoid the deficiencies of manually defined dependency relationships. At the same time, decompose the text into a word sequence, and use the BERT model to generate dynamic context word vectors to capture deep semantics. Subsequently, fuse semantic and syntactic information: construct a syntactic dependency graph based on word nodes and dependency edges, splice the BERT semantic embedding and syntactic relationship vectors to form an enhanced node representation, and then through the self-attention mechanism of the multi-layer graph attention network (GAT), dynamically calculate the weights between nodes and iteratively aggregate neighborhood features, and finally generate a node representation matrix that encodes both syntactic structure and semantic information, realizing the joint multi-feature modeling of legal texts.

[0009] Step 3: To further enhance the model's understanding of cross-sentence information, the present invention proposes a multi-hop reasoning module. This module uses a multi-scale graph aggregation strategy and position-aware graph encoding technology to construct a cross-sentence dependency relationship network, enabling the system to capture complex logical relationships hidden between multiple sentences.

[0010] This step includes three core parts:

[0011] Step 3.1 Syntax-guided multi-hop graph construction: Generate a syntactic dependency tree through syntactic analysis, and construct a cross-sentence anaphora resolution model to capture the logical relationships between sentences. Each sentence is connected to other sentences through its syntactic structure to form a global graph (G), which reflects the syntactic relationships and cross-sentence dependencies between sentences and provides a basis for subsequent multi-hop reasoning.

[0012] Step 3.2 Adaptive Graph Structure Enhancement Module: Based on the multi-hop graph, adjust the relationship strength between each sentence and other sentences through an attention-based gating mechanism. This module combines the structural information in the dependency syntax tree with a position-aware graph encoding strategy, automatically adjusts the feature contribution of the graph structure, and reduces noise propagation. Through adaptive gating, optimize the node features in the graph neural network, making the information more accurate in multi-hop propagation.

[0013] Step 3.3 Hierarchical Attention Aggregation Layer: Adopt a hierarchical attention mechanism to fuse multi-hop relationship information, combine local and global relationship information, and ensure that complex causal chains and logical inferences between sentences are captured from different subgraph levels. Through multi-head attention, dynamically fuse features at different levels, enhance the reasoning ability for key legal elements in case texts, and accurately infer violations in cases.

[0014] Step 4: Weighted Aggregation Stage: The model constructs a cross-attention weight mechanism through a sentence-level global relationship matrix, focuses on core sentences (such as key legal facts, clause citations, and case dispute focuses) with strong relevance to multiple violation categories, and dynamically adjusts the weights of each sentence embedding. Specifically, by calculating the association strength between sentences, weighted integration of the original sentence embeddings is performed, enabling sentences describing core elements such as violation behaviors and damage consequences to obtain higher weights, enhancing the model's ability to represent key information in the text, and thus improving the accuracy and interpretability of violation prediction.

[0015] Step 5: To further improve the prediction accuracy, the present invention combines a reinforcement learning strategy and uses a reward mechanism to optimize the model. By dynamically adjusting the feature extraction and reasoning processes in the model, ensure that the system can automatically adjust the prediction strategy according to the specific details of the case and improve the accuracy of information interaction over a long time span.

[0016] The present invention overcomes the difficulties in legal document parsing in traditional methods by combining multi-hop syntactic reasoning and semantic fusion technologies, and significantly improves the automatic matching ability of constituent elements and the accuracy of legal clause violation prediction. Through this method, a large number of case documents can be efficiently processed and analyzed, realizing automated violation detection and legal reasoning.

[0017] The beneficial effects of the embodiments of the present disclosure include:

[0018] (1) Multi-dimensional Feature Fusion Mechanism: This method first organically combines legal constituent element encoding, global semantic modeling, and syntactic logic analysis, overcoming the limitations of traditional methods that only rely on a single feature (such as semantic or rule matching). Capture global semantic features through deep context modeling in the semantic perception module, and explicitly construct a logical relationship network based on dependency analysis in the syntactic perception module. The two are effectively fused through a dynamic attention mechanism, providing strong support for capturing complex dependency relationships.

[0019] (2) Cross-sentence multi-hop reasoning ability: By introducing a multi-hop syntactic analysis module, a multi-hop syntactic dependency tree is constructed, successfully capturing the dependencies between sentences, thus significantly enhancing the ability to capture long-distance dependencies. Compared with traditional single-hop graph neural network models, this improvement enables the model to handle more complex legal logics and the associations between elements.

[0020] (3) Legal knowledge-guided optimization: The model transforms legal texts into structured legal constituent elements and aligns the prediction results with actual legal rules through a reinforcement learning framework. This optimization makes the model have a higher decision compliance when dealing with complex cases.

[0021] (4) Automated legal reasoning: A legal agent module is adopted to simulate expert-level legal reasoning and guide the reasoning process in combination with legal rules, enabling the model to more accurately identify and make violation predictions. Description of the Drawings

[0022] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0023] Figure 2 It is a model structure diagram corresponding to a method for predicting legal clause violations and dynamically matching constituent elements of the present invention; Detailed Embodiments

[0024] In order to enable those skilled in the art of the present technology to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion and highlight the core technical content of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0025] As Figure 1 shown, it is a schematic diagram of the process of a method for predicting semantic-syntactic fusion violations based on legal elements provided by an embodiment of the present disclosure, and the main process is as follows:

[0026] S1: Encoding of Legal Constituent Elements: First, for the input legal constituent elements, differential feature encoding is performed through a GRU group with parameter isolation to generate a basic feature vector. Then, a variance-driven dynamic screening mechanism is used to calibrate the feature vector and enhance the quality of feature representation.

[0027] S2: Semantic and Syntactic Fusion of Case Handling Event Documents: First, use the HanLP tool to perform syntactic dependency analysis on each sentence to extract the set of syntactic dependency paths. Then, use BERT to generate context-dependent word vector representations, and fuse semantic and syntactic relationships through the GAT self-attention mechanism to obtain enhanced word representations.

[0028] S3: Multi-hop Syntactic Analysis of Sentences in Case Handling Practice Documents: First, construct a syntax-guided multi-hop graph to capture cross-sentence logical relationships. Then, use an adaptive graph structure enhancement module to adjust feature contributions and reduce noise accumulation. Finally, fuse local and global relationship information through a hierarchical attention aggregation layer to achieve accurate legal reasoning.

[0029] S4: Weighted Aggregation of the Results of Semantic-Syntactic Fusion and Multi-hop Syntactic Reasoning: First, calculate the sentence-level global relationship matrix as the cross-attention weight to guide the model to focus on the core sentence set. Then, adjust the sentence embedding to ensure that key legal facts and clauses are accurately identified.

[0030] S5: Based on the encoding results of legal constituent elements and the weighted aggregation results after analyzing case handling event documents, legal agency is automatically matched and violation prediction is performed: First, calculate the correlation between the sentence representation and the LCE type embedding to generate a sentence representation specific to the LCE type. Then, the legal agency module automatically selects the subset of sentences with the highest discriminative power and performs iterative selection through a reinforcement learning agent. Finally, predict the violated legal articles and optimize the results by combining the reward mechanism for prediction accuracy and diversity control.

[0031] As Figure 2 shown, it is the system architecture diagram corresponding to a semantic-syntactic fusion violation prediction method based on legal elements provided by an embodiment of the present disclosure. The method mainly includes the following steps:

[0032] Step 1: Encoding of Legal Constituent Elements

[0033] Step 1.1 Input Legal Constituent Elements

[0034] For each accusation category g ∈ {1,..., G}, its corresponding LCE set is {LCE1,..., LCE K}. For the text sequence X of the k-th LCE (k) Perform differential feature encoding through a GRU group with parameter isolation to generate a basic feature vector

[0035] Step 1.2 Variance-driven Dynamic Screening Mechanism

[0036] Step 1.2.1 Calculate Dimension Features: For each dimension d, calculate the mean μ d and the variance Var(d).

[0037] Step 1.2.2 Generate Gating Weights: Obtain the gating weights g through a non-linear mapping (Sigmoid function) d , where when Var(d) → 0, g d → 1, enhancing the discriminability; when Var(d) is large, g d → 0, preserving the original distribution. g d The calculation process is as follows:

[0038] g d = 1 - σ(γ · Var(d)) (1)

[0039] Step 1.2.3 Calibrate Feature Vectors: Obtain the final feature vectors through gating redundancy filtering:

[0040]

[0041] Step 2: Semantic and Syntactic Fusion

[0042] Step 2.1 Syntactic Perception Analysis: Use the HanLP tool to perform syntactic dependency analysis on each sentence s i to generate a syntactic dependency tree and extract the set of dependency paths and identify key relationships (such as the subject-predicate relationship nsubj, verb-object relationship dobj, etc.).

[0043] Step 2.2 Semantic Perception Encoding: Decompose the legal document F into a word sequence W i = {w i,1 ,..., w i,m}, and for each word w i,j use BERT to generate context word vectors to fuse semantic and context information.

[0044] Step 2.3 Semantic-Syntactic Fusion

[0045] Step 2.3.1 Feature Fusion and Enhancement:

[0046] Construct a syntactic dependency graph G i = (V, E), where the nodes V are words and the edges E are dependency relationships By concatenating the semantic embeddings and syntactic relationship vectors obtain enhanced node representations These representations form the initial node matrix

[0047] Step 2.3.2 GAT Self-attention Mechanism:

[0048] The attention coefficient is calculated as follows, where represents the attention weight of node j to node k, dynamically allocating weights to different neighbor nodes.

[0049]

[0050] Through multiple-layer iteration (L times), GAT gradually aggregates node features and finally obtains the node representation matrix of the sentence:

[0051]

[0052] Step 3: Multi-hop Syntactic Analysis

[0053] Step 3.1 Syntactic-guided Multi-hop Graph Construction (SMGC)

[0054] For each sentence s i , generate a syntactic dependency tree and apply cross-sentence coreference resolution to construct a global graph G = (S, E). Among them, the node set is S = {s1, s2,..., s m}}, and the edge set is E = E coref ∪E syntactic (representing cross-sentence entity association and syntactic logical connection respectively).

[0055] Multi-hop Adjacency Tensor Construction:

[0056]

[0057] Association Strength Tensor Construction:

[0058]

[0059] Step 3.2 Adaptive Graph Structure Enhancement (AGSEM)

[0060] Step 3.2.1 Design a gating function for noise suppression, where G d dynamically adjusts the feature contribution of the d-th hop path and suppresses the noise edge weight.

[0061]

[0062] Step 3.2.2 Combine the adjacency matrix A d to update the edge weight to ensure that only the edge weights of valid paths are retained.

[0063] Step 3.2.3 Position-aware Bias: Introduce a bias term B ij= ρ(dist(s i , s j )) captures the hierarchical relationship between sentences, which captures the absolute position and relative structural information between sentences through the learnable function ρ.

[0064] Step 3.3 Hierarchical Attention Aggregation (HAA)

[0065] Step 3.3.1 Update multi-head attention, where Q (k) , K (k) , V (k) are the query, key, and value vectors of the k-th attention head. The gating modulation term and the position bias B guide the attention weights to balance the local (single-hop) and global (multi-hop) relationships. By fusing the semantic relationships at different subgraph levels through a hierarchical mechanism (such as "behavior description → legal provisions → consequence determination"), the causal chain in criminal charges is accurately captured.

[0066]

[0067] Step 3.3.2 Construct the global relationship matrix S global Quantifies the global association strength between sentences and provides a weight basis for subsequent weighted aggregation.

[0068]

[0069] Step 4: Weighted Aggregation

[0070] Step 4.1 Sentence Embedding Adjustment

[0071] In the weighted aggregation stage, the model uses the sentence-level global relationship matrix S global as the cross-attention weight to guide the model to focus on the set of core sentences highly relevant to multiple violation categories. These sentences usually contain key legal facts, legal provision citations, and the core issues of the case, which have a significant impact on charge prediction. Specifically, for the i-th sentence in the legal document, its adjusted sentence embedding f i is calculated by the following formula:

[0072]

[0073] where f j ' is the embedding vector of the original sentence j; represents the association strength between sentence i and sentence j, which is determined by the global relationship matrix S global . Through S global , the model can dynamically adjust the importance of sentences to ensure that key legal facts (such as behavior description, consequence determination, etc.) are assigned higher weights, thereby improving the accuracy of subsequent predictions.

[0074] Step 5: Automatic Matching and Prediction of Legal Representation

[0075] Step 5.1 Association of Sentences with LCE Types

[0076] The legal representation module generates sentence representations specific to LCE types by calculating the correlation between the sentence representation f i and the preset LCE type embedding e k :

[0077]

[0078] where denotes the concatenation or dot product operation, combining the sentence embedding with the LCE type embedding; f k,i is the enhanced representation of sentence i under LCE type k, ensuring its semantics is closely associated with the corresponding legal components (such as "subject", "object", etc.).

[0079] Step 5.2 Automatic Sentence Selection (Reinforcement Learning Agent)

[0080] The reinforcement learning agent iteratively selects the most discriminative sentence subset through the following steps:

[0081] Step 5.2.1 Aggregation of Selected Sentences

[0082] For each LCE type k of the input case document, the most critical features of the selected sentences are extracted through max pooling where w represents the feature dimension, and the maximum value across dimensions is retained to highlight the key information.

[0083] Step 5.2.2 Aggregation of Legal Elements

[0084] All LCE vectors e k,g generate an integrated representation through an attention-weighted mechanism where the weight is determined by the correlation between the sentence features and the LCE:

[0085]

[0086] where, W k is a learnable projection matrix, ensuring that the weight assignment matches the LCE type.

[0087] Step 5.2.3 Historical Embedding h t Tracks the agent's previous selections and is dynamically updated through the following formula:

[0088]

[0089] where, W pre and b preare trainable parameters that fuse historical selections with current features.

[0090] Step 5.2.4 The agent selects the next sentence based on the comprehensive score:

[0091]

[0092]

[0093] where are learnable interpolation weights that balance the importance of the LCE representation and the historical embedding h t ; W e and W h are trainable parameters that control the contributions of the LCE and historical information respectively. The agent selects the unselected sentences according to the probability to ensure the diversity and accuracy of the selection.

[0094] Step 5.2.5 The agent optimizes the selection strategy through the following formula:

[0095] r t = pred t - λ × rep t (16)

[0096] where pred t is the correctness of the prediction in the current step; rep t is the repetition penalty term that measures the similarity between the newly selected sentence and the historical selections; λ is the balance parameter that controls the weights of accuracy and diversity.

[0097] Step 5.3 Based on the highest discriminative power of the selected sentence, combined with the LCE type embedding e k,g , the final violation prediction is output Element encoding: Generate differentiated features through GRU and dynamic screening.

[0098] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A semantic-syntactic fusion violation prediction method based on legal elements, characterized in that, It includes the following steps: Step 1: Legal element encoding: Differentially encode the text sequence of legal elements through a group of gated recurrent units (GRUs) with parameter isolation to generate a basic feature vector; Step 2: Semantic-syntactic fusion modeling: Construct a syntactic dependency graph and fuse BERT semantic encoding, and aggregate node features through a graph attention network (GAT); Step 3: Multi-hop syntactic reasoning: Construct a cross-sentence multi-hop graph and implement adaptive graph enhancement, and capture the causal chain through a hierarchical attention mechanism; Step 4: Weighted aggregation: Generate a cross-attention weight matrix according to the global association strength between sentences, and perform weighted aggregation on sentence embeddings; Step 5: Automatic matching and prediction of legal agents: Combine legal element encoding and weighted features, and select key sentences through a reinforcement learning agent and output a violation prediction result.

2. The method according to claim 1, wherein The specific content of Step 1 includes: Step 1.1: For the text sequence of each legal element (LCE), perform differential feature encoding through a group of GRUs with parameter isolation to generate a basic feature vector; Step 1.2: Calibrate the feature vector based on a variance-driven dynamic screening mechanism, generate gating weights by calculating the mean and variance of dimensional features, and use the Sigmoid function to achieve redundant filtering.

3. The method according to claim 1, wherein The specific content of Step 2 includes: Step 2.1: Use the HanLP tool to construct a syntactic dependency tree and extract key dependency paths; Step 2.2: Generate context word vectors through the BERT model, and splice the semantic embedding and the syntactic relationship vector to construct an enhanced node representation; Step 2.3: Adopt the GAT self-attention mechanism to dynamically assign node weights, and form a node representation matrix through multi-layer iterative aggregation.

4. The method according to claim 3, wherein The specific content of Step 3 includes: Step 3.1: On the basis of Step 2.1, construct a global graph containing cross-sentence entity associations and syntactic logical connections through anaphora resolution, and use multi-hop adjacency tensors and association strength tensors to describe path relationships; Step 3.2: Suppress noise edge weights through a gating function, update the effective path edge weights in combination with the adjacency matrix, and introduce a position-aware bias to capture the absolute position and relative structure information between sentences; Step 3.3: Through the multi-head attention mechanism and the hierarchical mechanism, fuse the semantic relationships at different sub-graph levels to generate a global relationship matrix.

5. The method according to claim 4, wherein The specific content of Step 4 includes: Step 4.1: Use the global relationship matrix as the cross-attention weight to dynamically adjust the importance of sentence embeddings, where sentences corresponding to key legal facts are given higher weights; Step 4.2: Calculate the adjusted sentence embeddings, where the association strength is determined by the global relationship matrix.

6. The method according to claim 5, characterized in that The specific content of Step 5 includes: Step 5.1: Through the element encoding in Step 1 and the weighted features in Step 4, calculate the correlation between the LCE type embedding and the sentence representation to generate an enhanced representation specific to the LCE type; Step 5.2: Iteratively select sentences through a reinforcement learning agent, and extract key features and historical embedding tracking in combination with max pooling; Step 5.3: Adopt a formula to optimize the selection strategy, balance prediction accuracy and selection diversity, and output the final violation prediction.

7. The method according to claim 3, wherein The specific content of Step 2.3 includes: Construct a syntactic dependency graph with words as nodes and dependency relationships as edges, and generate enhanced node feature representations by concatenating semantic embedding vectors and syntactic relationship vectors; Form an initial node matrix with the enhanced node feature representations, and calculate the attention weights between nodes through GAT; Based on the attention weights, perform multi-layer iterative aggregation on the node features to generate a final node representation matrix that fuses syntax and semantics.

8. The method according to claim 4, wherein The specific content of step 3.2 includes: Step 3.2.1: Design a gating function for noise suppression, where the feature contribution of the d-th hop path is dynamically adjusted to suppress the noise edge weights; Step 3.2.2: Update the edge weights in combination with the adjacency matrix to ensure that only the edge weights of valid paths are retained; Step 3.2.3: Introduce a learnable position-aware bias term to capture the hierarchical relationships between sentences.

9. The method according to claim 6, wherein The specific content of step 5.2 includes: Aggregate the features of the selected sentences and extract key information through max pooling; Aggregate the attention-weighted representations of legal elements; Fuse the historical selection records and the current feature vectors to dynamically update the historical embeddings to track the agent's selections; Select unselected sentences according to the comprehensive scores, where the comprehensive scores include: (1) The prediction correctness weight; (2) A repetition penalty term used to quantify the similarity between the newly selected sentence and the historical selections; Coordinate the weights between prediction accuracy and selection diversity to optimize the selection strategy.

Images