A hierarchical search and graph reasoning-based vertical procedure intelligent analysis method and system

Abstract

The application belongs to the technical field of data processing, and specifically discloses a vertical regulation intelligent analysis method and system based on hierarchical retrieval and graph reasoning, wherein the method comprises the following steps: extracting a physical hierarchical topology of an unstructured regulation document, and mapping text fragments to a high-dimensional feature space to generate a text vector representation; based on the physical hierarchical topology and the text vector representation, identifying implicit logical associations between regulation clauses containing logical mutual exclusion associations, constructing directed edges with relationship weight coefficients to form a nonlinear regulation knowledge graph; based on the text vector representation, dynamically locking the retrieval boundaries of candidate text blocks by using the first derivative and the second derivative of the text feature semantic gradient; the application aims to solve the technical problems of semantic fragmentation caused by static hierarchical extraction in the vertical regulation analysis process, and the lack of bottom logic hard constraints in the generation stage of large language models, which easily produces analysis illusion.

Description

Technical Field

[0001] This invention relates to the technical field of information technology, specifically to a method and system for intelligent parsing of vertical category procedures based on hierarchical retrieval and graph reasoning. Background Technology

[0002] With the deepening of the global digital transformation wave, vertical industries such as energy and power, precision manufacturing, aerospace, and legal and government affairs have accumulated a massive amount of highly specialized procedures, standards, and normative documents. These vertical procedures, as core knowledge carriers guiding business operations, ensuring production safety, and maintaining industry standards, have profound strategic significance for improving industry production efficiency, reducing human error risks, and achieving precise reuse of knowledge assets through the application of intelligent analysis and automatic question-and-answer technologies.

[0003] In existing technological implementations, intelligent document parsing typically relies on semantic retrieval and machine reading comprehension technologies within deep learning frameworks. Traditional solutions often employ keyword matching or dense vector-based semantic retrieval mechanisms, pre-splitting procedural documents into fixed-length text blocks and mapping them to a high-dimensional vector space using a pre-trained language model.

[0004] However, existing technologies are increasingly revealing deep limitations when dealing with the special text type of vertical regulations. Specifically: First, existing hierarchical retrieval technologies, such as the intelligent bid parsing method based on multimodal knowledge graphs disclosed in Chinese patent application CN121413717A, typically employ static hierarchical division (document level, chapter level, etc.), lacking the ability to dynamically perceive semantic breaks in context, making them prone to misinterpretation. Second, existing graph reasoning retrieval methods, such as the semantic retrieval method for power grid regulations disclosed in Chinese patent CN120873208B, and the legal text information extraction method disclosed in CN120258001B, although constructing knowledge graphs that include hierarchical inheritance and responsibility collaboration and performing multi-hop path scoring retrieval, are essentially still at the retrieval level. When the retrieved conflicting clauses or implicit preconditions are fed into a large language model to generate answers, due to the lack of cross-modal hard constraints on the internal generation process of the large language model decoder, the model is prone to logical illusions, combining mutually exclusive regulations into self-contradictory analytical conclusions.

[0005] In summary, how to achieve mathematical-level dynamic physical boundary control in multi-granularity hierarchical retrieval, and how to implement hard constraint penalties on the underlying vocabulary of large language models through cross-modal mapping in the generation stage to completely eliminate logical illusions in procedure parsing, has become a pressing technical challenge in the field of knowledge intelligence in the industry. Summary of the Invention

[0006] This invention provides a method and system for intelligent parsing of vertical category procedures based on hierarchical retrieval and graph reasoning. The purpose is to solve the technical problems of semantic fragmentation caused by static hierarchical extraction during the parsing of vertical category procedures, and the fact that large language models are prone to parsing illusions due to the lack of underlying logical hard constraints during the generation stage.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for intelligent parsing of vertical procedures based on hierarchical retrieval and graph reasoning includes: extracting the physical hierarchical topology of unstructured procedure documents and mapping text fragments to a high-dimensional feature space to generate text vector representations; based on the physical hierarchical topology and text vector representations, identifying implicit logical associations between procedure clauses that contain mutually exclusive logical relationships, and constructing directed edges with relational weight coefficients to form a nonlinear procedure knowledge graph; based on the text vector representations, dynamically locking the retrieval boundaries of candidate text blocks using the first and second derivatives of the semantic gradient of text features, and reducing the dimensionality of the interaction feature matrix between the query vector within the boundary and the candidate text blocks to... The matching score is used to output the retrieval primitive; using the retrieval primitive as the anchor point, a local correlation subgraph is extracted from the nonlinear procedural knowledge graph, and a feature aggregation and discarding operation is performed on the low-priority node information with logically mutually exclusive correlation edges using a dynamically generated zero gating mask coefficient, outputting a set of logical topological nodes that resolve mutual exclusion contradictions; the retrieval primitive and the set of logical topological nodes are input into a deep large language model, the low-priority node information that has been feature aggregated and discarded is transformed into a sparse penalty mask vector of the vocabulary dimension, and added to the original log probability vector to force the generation probability of the illegal words to be set to zero, producing an analytical conclusion.

[0008] In one aspect of the present invention, the process of extracting the physical hierarchy topology of unstructured procedure documents and mapping text fragments to a high-dimensional feature space to generate text vector representations of the extracted physical hierarchy topology specifically includes: A hierarchical prediction function is established using a combination of hierarchical gated recurrent units and conditional random fields. The physical hierarchy topology is extracted by assigning corresponding weight coefficients to the layout feature functions corresponding to each layout feature and performing a summation operation.

[0009] In one aspect of the present invention, the process of identifying implicit logical associations containing mutually exclusive logical relationships between procedural clauses based on the physical hierarchy topology and text vector representation, and constructing directed edges with relational weight coefficients to form a nonlinear procedural knowledge graph, specifically includes: Utilize a hard constraint recognizer based on regular expressions to capture explicit reference relationships between procedure clauses; By using a deep learning-based weak association classifier, we can identify semantically overlapping logical implicit relationships by calculating the divergence of entity distributions among different clauses.

[0010] In one aspect of the present invention, the process of dynamically locking the retrieval boundary of candidate text blocks based on the text vector representation, utilizing the first and second derivatives of the semantic gradient of the text features, and reducing the dimensionality of the interaction feature matrix between the query vector and the candidate text blocks within the boundary to a matching score to output the retrieval primitive, specifically includes: The dot product of the query vector and the transpose of the candidate text block key feature matrix is ​​normalized using a normalized exponential function; The interaction feature matrix is ​​obtained by multiplying the normalized result by the value matrix; Perform mean pooling on the interaction feature matrix, multiply the pooling result by the linear projection weight matrix and add a bias term to output the matching score in one-dimensional scalar form.

[0011] In one aspect of the present invention, the process of extracting local correlation subgraphs from the nonlinear procedural knowledge graph using the retrieval primitive as an anchor point, and performing feature aggregation and discarding of low-priority node information with logically mutually exclusive correlation edges using a dynamically generated zero gating shielding coefficient, and outputting a set of logically topological nodes that resolve mutual exclusion contradictions, specifically includes: Perform a scalar multiplication operation between the gated shielding coefficient and the corresponding neighbor node features; Perform mean aggregation on the features after multiplication to obtain the aggregated neighbor feature vector; The aggregated neighbor feature vectors are concatenated with the feature vectors of the previous layer nodes, and the updated node features are obtained by multiplying them by a weight matrix and a nonlinear activation function in sequence.

[0012] In one aspect of the present invention, the process of inputting the retrieval primitives and the logical topology node set into a deep large language model, converting the low-priority node information discarded by feature aggregation into a vocabulary-dimensional sparse penalty mask vector, and adding it to the original log-probability vector to force the generation probability of violating words to be set to zero, and producing an analytical conclusion, specifically includes: For illegal words that are strongly related to the low-priority node information, the element value corresponding to it in the sparse penalty mask vector of the word list dimension is set to a minimum value that tends to negative infinity, so that the generation probability after addition and normalization exponential function processing approaches zero. For words that are unrelated to the low-priority node information, the element values ​​in the sparse penalty mask vector of the vocabulary dimension are kept to zero.

[0013] In one aspect of the invention, the method further includes: The deep large language model simultaneously outputs a visual derivation path corresponding to the parsing conclusion. The visual derivation path is presented in the form of a directed acyclic graph, where nodes represent procedural clauses and edges represent reasoning logic.

[0014] In another aspect, the present invention also relates to a vertical category procedure intelligent parsing system based on hierarchical retrieval and graph reasoning, for implementing the above method, the system comprising: The document structuring module is used to extract the physical hierarchical topology of unstructured procedure documents and map text fragments to a high-dimensional feature space to generate text vector representations. The knowledge graph construction module is used to identify implicit logical associations containing mutually exclusive logical associations between procedure clauses based on the physical hierarchy topology and the text vector representation, and to construct directed edges with relation weight coefficients to form a nonlinear procedure knowledge graph. The dynamic hierarchical dual-path retrieval module is used to dynamically lock the retrieval boundary based on the text vector representation, using the first and second derivatives of the semantic gradient of the text features, and reduce the dimensionality of the interaction feature matrix within the boundary to the matching score to output the retrieval primitive. The conflict resolution graph reasoning module is used to extract local correlation subgraphs in the nonlinear procedure knowledge graph using the retrieval primitives as anchors. It uses a dynamically generated zero gating shielding coefficient to perform feature aggregation and discarding on low-priority node information with logically mutually exclusive correlation edges, and outputs a set of logical topological nodes that resolve mutual exclusion contradictions. The structured intelligent parsing generation module is used to input the retrieval primitives and the logical topology node set into a deep large language model, convert the low-priority node information that is discarded by feature aggregation into a sparse penalty mask vector of the vocabulary dimension, and add it to the original log probability vector to force the generation probability of the illegal words to be set to zero, thus producing a parsing conclusion.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Firstly, in the multi-granularity hierarchical retrieval stage, this invention overcomes the technical bottleneck of traditional fixed-step slicing, which easily leads to semantic misinterpretation. It introduces the first and second derivatives of the semantic gradient of text features to dynamically lock the retrieval boundary, uses a calculus mechanism to accurately extract text blocks containing complete context, and reduces the dimensionality of the interaction feature matrix to a one-dimensional scalar form of matching score within the retrieval boundary, which greatly improves the retrieval recall completeness and ranking accuracy of complex cross-chapter length dependent information.

[0016] Secondly, in the graph reasoning and conflict resolution stage, in response to the hidden logical mutual exclusion relationships in the regulations, this invention does not rely on the natural understanding of the generative model, but instead dynamically generates a zero gating shielding coefficient to perform feature aggregation and discarding on low-priority nodes, thus completing the hard isolation and resolution of mutual exclusion contradictions at the root of topological computation in the spatial domain graph neural network.

[0017] Most importantly, in the intelligent parsing generation stage, this invention constructs a cross-modal decoding intervention mechanism. This mechanism precisely transforms low-priority node information discarded during feature aggregation into a sparse penalty mask vector at the vocabulary dimension, and directly adds it to the original log-probability vector output by the deep large language model. This underlying tensor-level operation forcibly sets the generation probability of violating graph mutual exclusion constraints to zero, ensuring the smooth generation of irrelevant words while completely cutting off the path for large language models to generate parsing illusions. In summary, this invention, through the reconstruction of the underlying mechanism, achieves a generational leap in parsing accuracy and an extreme reduction in the occurrence rate of illusions by sacrificing millisecond-level inference computation. This provides a highly deterministic and traceable technological foundation for vertical industries such as power and law, which have zero tolerance for logical fallacies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a vertical category procedure intelligent parsing method based on hierarchical retrieval and graph reasoning according to the present invention.

[0020] Figure 2 This is a framework diagram of an intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning, according to the present invention.

[0021] Figure 3 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0023] Example 1 Please see Figures 1-3As shown, this embodiment discloses a method and system for intelligent parsing of vertical category procedures based on hierarchical retrieval and graph reasoning, wherein the method specifically includes: The first step involves deep structured parsing and multi-dimensional feature space mapping of vertical-category procedural documents. The system uses a pre-defined structure recognition model to perform hierarchical parsing of the original unstructured procedural documents, extracting the document's physical hierarchical topology. This topology includes a tree-like structure with four levels of interconnected chapters, sections, subsections, and subheadings. A hierarchical prediction function is established to weight the document's formatting, indentation features, and specific trigger words. The calculation formula is as follows:

[0024] in, Slice the input text; This represents the total number of preset layout features; For the first Item typesetting feature functions are used to extract text fragments. Extract specific feature values ​​such as font size, bolding, indentation distance, and numbering pattern; In order to be with the first The weight coefficients corresponding to the item layout features, which can be learned through training; The output is a hierarchical prediction value used to ultimately determine which level of chapter, section, article, or category the text segment belongs to.

[0025] During the parsing process, a sequence labeling algorithm based on a bidirectional encoder representation transformer is used to identify and extract metadata information for each level of node. For each text segment, a dense vector encoder is used to map it to a high-dimensional feature space to generate a text vector representation.

[0026] The second step is to construct a procedural knowledge graph based on logical topology. Through semantic relation extraction operators, implicit logical relationships between procedural clauses are identified. These logical relationships include referencing relationships, inclusion relationships, pre-constraint relationships, and logically mutually exclusive relationships. For each pair of related clause nodes... and Establish directed edges And assign corresponding relation weight coefficients. The relation weight coefficient The calculation formula is: in, For nodes With nodes The relationship weight coefficient For nodes With nodes Normalized values ​​of the correlation frequencies between them For nodes With nodes Semantic similarity between them For nodes With nodes The hierarchical distance between them It is the reciprocal of the hierarchical distance. , , These are the corresponding weight hyperparameters. Ultimately, this forms a non-linear knowledge graph with procedural clauses as vertices and logical relationships as edges.

[0027] The third step involves performing a multi-granularity hierarchical retrieval with a dynamic boundary adjustment mechanism. Upon receiving a query request, the system filters out the target chapter. Within the target chapter, candidate text blocks are extracted using a sliding window, and the derivative of the semantic similarity reduction rate of adjacent text blocks is calculated. As the sliding window expands to the physical neighborhood, the semantic gradient of the text features of adjacent windows is calculated in real time. If the second derivative of the semantic gradient reaches an inflection point, or the first derivative falls below a preset continuity threshold, the expansion is dynamically stopped, thus determining the retrieval boundary containing the complete context. Within the determined boundary, a cross-attention mechanism is used to calculate the interaction feature matrix between the query vector and the candidate text block, and this matrix is ​​then reduced to a one-dimensional scalar matching score through linear projection and pooling operations. The calculation formula is as follows:

[0028]

[0029] in, This is the cross-attention feature matrix. For query vector, The key feature matrix, For value matrices, For feature dimension, To match scores, The learnable dimensionality-reduced projective weight matrix. It is the mean pooling function. This is a bias term. Reorder the candidate text blocks and output the search primitives.

[0030] The fourth step involves implementing logical path reasoning based on subgraph sampling and conflict resolution. Using the node to which the retrieval primitive belongs as the anchor point, multi-hop neighborhood expansion is performed in the procedure knowledge graph to extract locally related subgraphs. Spatial domain graph neural networks (GraphSAGE architecture) are used to aggregate node features within the subgraph. Before feature aggregation, the system introduces an isolation mechanism based on priority: when adjacent nodes with "logically mutually exclusive association" edges are identified, according to a preset priority principle (e.g., specialized rules have greater effect than general rules), the feature vectors of low-priority nodes are negatively weighted or their features are directly discarded during aggregation. The updated node feature calculation formula is as follows:

[0031] in, As an aggregation feature, Features of neighboring nodes The gating shielding coefficient (0 or 1) is dynamically generated based on the effectiveness priority. For splicing operations, This is the weight matrix. This is the activation function. The graph reasoning engine identifies relevant implicit terms and preconditions and outputs a set of logical topological nodes whose mutual exclusion contradictions have been resolved.

[0032] The fifth step involves executing structured intelligent parsing and generation with a sparse penalty masking mechanism. The retrieved primitives and the features of mutually exclusive topological nodes that have been resolved are input into the deep large language model. During the decoder generation stage, a mandatory hard constraint mechanism is constructed to map "graph topological nodes" across modalities to the "large model autoregressive vocabulary." Specifically, the system pre-constructs a "procedure node-vocabulary association mapping matrix," transforming the graph node information judged as "mutually exclusive and low-priority (should be excluded)" in the fourth step into a vocabulary-dimensional penalty mask vector in the large model output space. In the cross-attention layer of the decoder, the query vector of the large model simultaneously queries the key matrix of the text context and the key matrix of the graph node features, achieving orthogonal fusion of features. In calculating the... Each decoding time step generates candidate words. When calculating the probability distribution, the formula is used:

[0033] in, For generating words The probability distribution, The original log-odds vector (Logits) output by the large language model. Let be the sparse penalty mask vector. For associated words that violate the graph mutual exclusion constraint, its in The corresponding element value is set to (Tending towards negative infinity), the generation probability of related illegal words is forcibly set to zero; for irrelevant words, The element values ​​are kept at 0 to ensure that the fluency of the generated language is not compromised. The system's final output strictly follows the analytical conclusions and visual derivation paths of the procedural logic topology.

[0034] In a preferred embodiment of the present invention, in the first step, the structure recognition model adopts a combined architecture of hierarchical gated recurrent units and conditional random fields. In a preferred embodiment of the present invention, in the second step, the semantic relation extraction operator includes a hard-constraint recognizer based on regular expressions and a weak-association classifier based on deep learning. In a preferred embodiment of the present invention, in the fifth step, the logical derivation path is presented in the form of a directed acyclic graph, achieving transparency and traceability of the parsing process.

[0035] This invention also provides a vertical-category procedure intelligent parsing system based on hierarchical retrieval and graph reasoning. The system includes: a document structuring module for receiving procedure documents, extracting physical hierarchical topology, and mapping text to semantic vectors; a knowledge graph construction module for establishing a nonlinear procedure knowledge graph containing references, constraints, and mutually exclusive associations, including a graph database storage unit; a dynamic hierarchical dual-path retrieval module, configured with a semantic gradient calculation unit and a vector dimensionality reduction projection unit, for performing dynamic hierarchical retrieval based on first / second derivative calculus control boundaries and calculating a one-dimensional scalar matching score; and a conflict resolution graph reasoning module, configured with a graph neural network processor and an effectiveness gating unit, for performing subgraph mining and using a masking coefficient. Complete the topology resolution of mutually exclusive logical nodes. The structured intelligent parsing generation module, with an embedded large language model decoder containing sparse penalty mask mapping, is used to... The formula performs vocabulary-level generation interception on output words that violate logical topology, producing analytical conclusions and visual derivation paths.

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0037] In some optional embodiments, the specific usage method is as follows: Step 1: Perform deep structured parsing and multidimensional feature space mapping of vertical category procedure documents; Step 1-1: Visual Layout Analysis and Physical Region Segmentation. A deep learning-based layout analysis algorithm (computer vision model) is invoked to perform visual-level pixel segmentation of the document page, forcibly separating the title line, body text blocks, tables, images, and header / footer areas to obtain clean physical coordinate information.

[0038] Steps 1-2: Physical hierarchy topology extraction and metadata annotation. Text is extracted using Optical Character Recognition (OCR) technology. The system constructs a hierarchy prediction function. This includes typesetting features such as font size, bolding, indentation, and specific numbering patterns. Weighted calculations are performed to absolutely reconstruct the four-level tree topology of "chapter, section, article, and subsection". Then, a sequence labeling algorithm based on a bidirectional encoding representation transformer is run to extract the hierarchical index, title / content summary vectors, and constraint attribute labels.

[0039] Steps 1-3: Multidimensional Feature Mapping and Vector Representation Generation. The smallest text segment is input into a dense vector encoder, mapped to a semantic vector representation in a high-dimensional feature space. Detailed Explanation of Core Terms and Algorithms: Hierarchical Gated Recurrent Unit and Conditional Random Field (GRU+CRF): The core architecture of sequence labeling. GRU is responsible for capturing the contextual memory state in long texts, while CRF is responsible for outputting the globally optimal legal hierarchical transition path, ensuring the physical accuracy of hierarchical parsing. Domain-Specific Contrastive Learning Strategy: A specialized training scheme for dense vector encoders. The algorithm forcibly constructs negative sampling pairs of "procedure terms - corresponding business scenarios." Through this mechanism, the model is forced to distinguish highly similar industry-specific terms in the feature space.

[0040] Specifically, in the analysis of 52 documents totaling approximately 1.2 million words from a large power company's "Compilation of Safety Production Regulations," the document structuring module accurately restored the physical boundaries of 4,520 third-level headings and 12,800 fourth-level headings. The system employs a pre-trained BERT encoder, strictly setting the output vector dimension to 768 dimensions. By injecting domain features through a contrastive learning strategy, for easily confused power-related terms such as grounding and lap joints, the encoder's Euclidean distance in the 768-dimensional high-dimensional space is forcibly increased by more than 45%, fundamentally eliminating the semantic ambiguity drawbacks of general NLP models.

[0041] Step 2: Constructing a procedural knowledge graph based on logical topology; 2-1: Extraction of Explicit and Implicit Semantic Relationships. The semantic relationship extraction operator is invoked to perform a dimensionality reduction scan on the clause text. Explicit logic is extracted using a hard constraint recognizer, and implicit logic is extracted using a weak association classifier.

[0042] Step 2-2: Topology network weight calculation and graph database instantiation. Instantiate the procedure clauses as a set of vertices. Logical relationships are instantiated as a set of directed edges. For any associated node and Establish directed edges And use formulas to assign relationship weight coefficients. The weights are determined by association frequency, semantic similarity, and physical hierarchical distance, ultimately generating a nonlinear knowledge network. And persist it to the graph database.

[0043] Among them, the hard constraint recognizer and the weak association classifier refer to the following: the hard constraint recognizer is based on regular expressions and forcibly captures, such as the deterministic explicit references that conform to the XX standard (see Article X). The weak association classifier is based on deep learning and calculates the KL divergence of entity distributions between different clauses. When the divergence approaches a minimum value, the system determines that there is a logical implicit relationship with highly overlapping semantics.

[0044] Domain ontology constraint and dependency syntax tree analysis refers to predefining core entities and standard relation lexicons in a vertical scenario. The system accurately captures the implicit agent and patient logic by analyzing the subject-verb-object dependency syntax tree of sentences. For example, if clause A stipulates that it must satisfy a core parameter of clause B, the algorithm immediately generates a constraint support path from B to A.

[0045] Specifically, in the topology construction of the aforementioned power dataset, the extraction module cumulatively performed instantiation operations, generating a total of 15,600 explicitly referencing directed edges and 8,400 implicitly related edges based on KL divergence. This was followed by the calculation of global relation weight coefficients. At that time, the system's hyperparameter was set as: correlation frequency weight. Semantic similarity weight Hierarchical distance weight The topology graph constructed using this parameter allocation ensures the convergence of subsequent multi-hop inference.

[0046] Step 3: Perform multi-granularity hierarchical retrieval; Step 3-1: Macro-level global positioning driven by inverted index. Receive query requests. Then, using a global retrieval operator, mixed matching is performed at the chapter and section levels to extract the most semantically relevant results. One target chapter.

[0047] Step 3-2: Sliding window mechanism and cross-attention micro-matching. In Within each target chapter, a sliding window mechanism is activated to traverse text blocks. The query vector and candidate text blocks are fed into a cross-attention network to calculate the interaction feature matrix between the query and the text blocks, and then reduced to a one-dimensional scalar matching score. The specific calculation formula is as follows:

[0048]

[0049] in, This is the cross-attention feature matrix. For query vector, The key feature matrix of the candidate text blocks. This is the value matrix of the candidate text blocks. For feature dimension, It is the mean pooling function. and These are the learnable dimensionality-reduced projective weight matrix and the bias term, respectively. This is the final match score.

[0050] Step 3-3: Dynamic Boundary Adjustment and Semantic Completion. Monitor candidate block scores and trigger continuity verification. When the score exceeds a preset threshold... At this point, the algorithm forcibly expands the boundary bidirectionally to the physical neighborhood until the first derivative of the relevance (the rate of decline) exceeds the circuit breaker ratio, thereby locking in the complete retrieval primitive.

[0051] The hybrid retrieval of inverted index and vector space model refers to a two-stage recall process in the macro-positioning stage. First, the BM25 algorithm is run for hard matching recall of the inverted index to ensure that core professional vocabulary is not lost. Then, a domain-pre-trained language model is used to calculate the high-dimensional cosine similarity between the query and the document to complete the secondary fine ranking.

[0052] The topic clustering logical hierarchical (extended architecture) refers to the system's introduction of the implicit Dirichlet distribution algorithm for full-scale topic modeling, targeting massive amounts of entries (such as a national-level legal database). A logically sharded index based on topic clustering is constructed, and during retrieval, topic routing is triggered first to reduce the computational power consumption range, achieving millisecond-level response times.

[0053] Specifically, the input query request is: ventilation requirements during the maintenance of sulfur hexafluoride (SF6) equipment in a substation. The hybrid retrieval system first identifies the target interval as the two chapters on "General Requirements for Substation Operation and Maintenance" in the first-level macro-routing. Then, the lower-level system activates a sliding window to extract candidate blocks, strictly limiting the window size to 512 tokens and the step size to 128 tokens. Through cross-attention network calculations, the system accurately segments five micro-text fragments with confidence scores strictly greater than 0.85, using these as absolute anchor points for the next step of graph sampling.

[0054] Step 4: Implement the logical path reasoning step based on subgraph sampling; 4-1: Multi-hop neighborhood sampling of knowledge graphs. Using the data extracted in step 3... Using high-confidence text nodes as absolute anchors in the underlying graph Perform fixed-length multi-hop expansion along the directed edge in the middle, and extract the local correlation subgraph containing the preconditions. .

[0055] Step 4-2: Drive the graph neural network processor to perform computations on the subgraph. For example, a graph convolutional network (GCN) can be used for feature propagation and aggregation.

[0056] in, For the first Layer node characteristics, It is an adjacency matrix with self-connections. For based on The calculated degree matrix, For the first The trainable weight matrix of the layer, This is the activation function.

[0057] Step 4-3: Consistency Detection and Conflict Resolution. In the feature aggregation process, to resolve logically mutually exclusive contradictions, this invention introduces a gating and shielding mechanism based on validity priority. This embodiment uses the GraphSAGE architecture for neighbor feature aggregation with conflict resolution. The node feature update process is as follows:

[0058]

[0059] MEAN(...) performs mean aggregation on the set of neighboring nodes.

[0060] {.}: Defines a set of neighboring nodes. The parentheses (.) enclose the entire set.

[0061] Neighboring nodes In terms of the characteristics of the upper layer, For nodes The neighborhood group, The gating shielding coefficient is dynamically generated based on the effectiveness priority (when an edge is identified). Logically mutually exclusive and nodes Priority lower than hour, ,otherwise ), The aggregated neighbor characteristics, This is a vector concatenation operation. This is the weight matrix. For activation function, For the updated node Its characteristics.

[0062] The attention-based subgraph sampling algorithm refers to the inference engine refusing to traverse the entire graph to prevent memory overflow. In addition to sampling first-order paths, the algorithm also runs a path sorting algorithm to filter high-weight second- and third-order paths. The network structure adopts a three-layer GraphSAGE architecture, with the aggregation function set to mean aggregation, and the number of hidden units in each layer fixed at 512, 256, and 128. The sampling strategy uses random walks, with a hard constraint: the maximum depth of a single inference attempt does not exceed 3 hops.

[0063] The consistency check and validity priority principle refers to the core operators for avoiding legal risks. When contradictory clauses are detected, the algorithm refuses to output them directly. Instead, it applies objective validity weights, such as "detailed rules take precedence over general rules," to eliminate low-weight paths and retain only legitimate reasoning branches.

[0064] Specifically, the graph reasoning engine inputs the five maintenance and ventilation text fragments identified in the preliminary steps into the graph. The system performs a maximum of three-hop neighborhood expansion. Through graph convolution calculations, the algorithm overcomes the limitations of surface text and successfully locates two mandatory supporting clauses in a completely independent section of labor protection management regulations. Furthermore, when testing conflict queries regarding work approval permissions under special weather conditions, the subgraph detects mutually exclusive logical edges between the general rules and detailed rules. The consistency operator immediately determines the conflict and automatically executes the resolution instruction that prioritizes the detailed rules over the general rules, directly truncating the erroneous approval path output.

[0065] Step 5: Intelligent parsing and generation steps that integrate multi-source information; 5-1: Multimodal Tensor Fusion and Decoding Input. The primitive tensors obtained from hierarchical retrieval, the implicit clause tensors of the GNN inference path, and the subgraph feature vectors are aligned and uniformly fed into the Transformer decoder with external knowledge injection.

[0066] Step 5-2: Constrained Generation and Logical Constraint Mechanism. When the decoder predicts sequence probabilities, a logical constraint attention layer is activated. The system performs intervention calculations when generating each token to ensure that the direction of text generation is strictly constrained to the logical topology produced by graph reasoning.

[0067] Step 5-3: Structured analysis results and source map output. The model not only outputs natural language conclusions, but is also required to output directed acyclic graph (DAG) paths with white-box properties, and explicitly label all entity references supporting the conclusions.

[0068] Among them, the logical constraint attention layer and the probability cliff penalty refer to the fundamental mathematical means to completely eliminate the "illusion" phenomenon of large models. Hard constraints are achieved by modifying the probability distribution of the Softmax function. When generating the... Word At that time, the system reverse-checks the subgraph state vector. If a predicted word is detected to contradict a logical chain node, the system immediately triggers a probability penalty, setting its generation probability directly to near zero.

[0069] The derivation and visualization of a Directed Acyclic Graph (DAG) refers to the dimensionality reduction mapping of the black-boxed neural network output into a human-readable topological graph. Specific terms for graph node mapping, and constraints and referential relationships for directed edges, thoroughly meet the stringent requirements of vertical industries for system reliability and 100% traceability.

[0070] Specifically, the entire intelligent parsing system is deployed on a high-performance distributed computing cluster. The text vector library is indexed using the HNSW (Hierarchical Navigable Small World) algorithm, ensuring millisecond-level nearest neighbor search. After rigorous processing through all five steps, the final output of the system eliminates probabilistic patchwork. Experimental data shows that by forcibly adding graph feature attention heads and Softmax probability penalties, the system's parsing illusion rate significantly decreased from 14.8% in the conventional RAG architecture. The accuracy of structured parsing was greatly improved, and the logical derivation consistency score increased from 68.2% to 95.8%, at the cost of only increasing the response time for a single complex cross-chapter query to 2.4 seconds. This is a highly commercially viable technical implementation path that sacrifices a small amount of computing power for absolute logical determinism.

[0071] In this embodiment, the "Compilation of Safety Production Regulations" of a large power company was selected as the experimental dataset. This compilation contains 52 professional regulation documents, totaling approximately 1.2 million words, covering multiple fields such as substation and power transmission. The document structure is complex, containing a large number of nested lists, cross-chapter references, and mutually exclusive constraints (such as conflicts between general regulations and special operation approval rules).

[0072] During the operation, the document structuring module is first used to parse the procedures, identify level 3 and level 4 headings, and complete dense vector encoding (768 dimensions). In the knowledge graph construction module stage, explicit and implicit connection edges are extracted, and the hyperparameter of the relationship weight coefficient is set as follows: , , In the retrieval phase, the query request is set to the approval authority for substation off-site operations under severe weather conditions. The dynamic hierarchical dual-path retrieval module initiates semantic gradient monitoring when extracting candidate blocks. When the sliding window encounters a sudden content change (jumping from the approval process to equipment parameter descriptions) during outward expansion, the first derivative of the semantic gradient drops sharply, and the system immediately triggers a hard stop, accurately capturing the complete context of operations under severe weather conditions and eliminating semantic fragmentation caused by traditional static segmentation. When the conflict resolution graph reasoning module performs GraphSAGE multi-hop expansion, it discovers a mutually exclusive edge between the general principle node (station manager approval) and the detailed principle node (requiring special approval from the safety supervision department for rain and snow). The module triggers the effectiveness gating unit, determining that the effectiveness of the special weather detailed principle is higher than that of the general principle, and then aggregates the masking coefficient of the general principle node. Setting it to 0 successfully resolved the rule conflict at the topology level. During the parsing and generation phase, the structured intelligent parsing and generation module receives the Logits output from the large language model. Since the station manager approval node is determined to be invalid and mutually exclusive in graph reasoning, the system hits the lexical units such as "station manager" and "duty manager" in the procedure node-lexical association mapping matrix, and at time step... When generating these candidate words, they are placed in the sparse penalty mask vector. The corresponding value in is set to This cross-modal, low-level intervention ensures that the large model can never output the wrong answer approved by the station manager, completely cutting off the path to generating analytical illusions, and ultimately outputting the correct structured result specially approved by the safety supervision department.

[0073] The weight hyperparameters involved in the above embodiments of the present invention Preset continuity threshold, circuit breaker ratio, preset threshold and dynamic threshold The values ​​are all derived from a limited number of historical business data points, obtained through a series of data processing steps including grid search and multiple rounds of cross-validation. Specifically, the system first extracts 10,000 valid retrieval and inference records from various historical procedures as training samples, setting the target loss function as the harmonic average of parsing precision and recall; then, iterative calculations are performed within the parameter space using an automated hyperparameter optimization algorithm, with a preset threshold. Taking the value of as an example, The search range is set to 0.6 to 0.95, with a step size of 0.05. When the value is 0.80, the model achieves an average matching accuracy of 85% on the test set, while when... When the value was increased to 0.85, the matching accuracy reached 92% and the recall rate did not decrease significantly; when When the threshold was increased further to 0.90, the recall rate plummeted due to overly stringent boundary conditions. Therefore, the preset threshold was ultimately adjusted. The best practice value is locked at 0.85; all other preset parameters and thresholds of the system are calculated using this data-driven optimization mechanism and evaluation process.

[0074] Example 2 Please see Figure 3 As shown, this embodiment is a further optimization based on embodiment one. In this embodiment, in order to solve the problem of rampant topological noise caused by static graph reasoning in the prior art, this embodiment transforms the edge weights of the graph network from static prior to intent posterior, realizing the deep algorithm fusion of query intent and layer-level topology. In the embodiment, during the implementation of subgraph sampling-based logical path reasoning in S4, the system introduces a query intent-driven dynamic topology weighting algorithm. The algorithm first constructs a query intent representation model to extract the query intent vector, and then uses a dynamic weight calculation formula to perform tensor concatenation and fusion activation of the query intent vector and the original edge features in the static logical graph to calculate a dynamic edge weight matrix adjusted by intent. The dynamic edge weight matrix directly replaces the original adjacency matrix and is input into the graph convolutional neural network to achieve real-time reshaping of the topology of the local associated subgraph.

[0075] After reshaping the local relational subgraph, the system executes an adaptive subgraph pruning mechanism. This mechanism uses the isolated forest algorithm model to detect outliers by using the weighted degree centrality of nodes under the dynamic edge weight matrix as a feature dimension. Redundant nodes with weighted degree centrality lower than the dynamic threshold are forcibly pruned, thereby generating a simplified logical subgraph with high intent relevance. The node feature aggregation result of the simplified logical subgraph is then output to the intelligent parsing and generation module.

[0076] Specifically, this embodiment abandons conventional static reasoning and designs an interactive and collaborative model of intent and topology. The smallest implementable unit for algorithm fusion is as follows: Step 1. Query Intent Representation Model Construction: Extracting User Queries The core intent features.

[0077]

[0078] in, The output is the query intent tensor. For pre-trained language model encoders, The intention feature projection matrix, For bias vectors, This is a layer normalization operation.

[0079] Step 2. Dynamic Topology Edge Weight Generation Model (interacting with upstream algorithm): Receives the output of the preceding formula. , and the static edge feature vectors generated during the graph construction phase Then, the parts are spliced ​​and merged.

[0080]

[0081] in, For nodes With nodes The dynamic edge weight scalar between them This indicates a tensor splicing operation. To fuse the weight matrix, To incorporate the bias term, This is the Sigmoid activation function.

[0082] Step 3. Dynamic Subgraph Convolutional Inference Model (Interacting with Upstream Algorithm): Outputting the pre-formula... Inject adjacency matrix to construct dynamic adjacency matrix The update process of a graph neural network becomes:

[0083] in, To incorporate the adjacency matrix with dynamic edge weights, For based on The calculated dynamic degree matrix, For the first Layer node feature representation (at layer 0, this vector inherits the text vector representation output from the layered retrieval stage). For the first Learnable convolutional kernel parameters for each layer It is a non-linear activation function.

[0084] In practical applications, a contrastive learning loss function is used for joint training of the model. Positive sample pairs represent the core supporting clauses of historical query statements with expert annotations, while negative sample pairs represent irrelevant paths generated by random walks. The model is updated via gradient descent and backpropagation. , and Training parameters: learning rate set to 0.0005, batch size set to 64, and AdamW optimizer used.

[0085] During model application, in the real-time query phase, the system first calculates dynamic edge weights and then performs adaptive pruning. It enumerates the weighted degree centrality of all nodes (including absolute degree, relative degree, out-degree, and in-degree). If the sum of the dynamic in-degrees of a node is lower than a set dynamic threshold... If it is determined to be an intention-irrelevant node, it will be forced to leave the subgraph. Remove from the list.

[0086] In addition, this method also needs to use a continuously differentiable penalty function to directly map the graph topological features of the direction-1 output to the probability distribution adjustment factor of the decoder, so as to achieve deep collaboration between the bottom graph algorithm and the top generation algorithm. The specific steps are as follows: In the fifth step of intelligent parsing and generation that integrates multi-source information, the system constructs an adaptive decoding penalty model that integrates topological features. During the decoding process at each time step, the model calculates the Euclidean distance between the current hidden state vector of the decoder and the feature vectors of all core constraint nodes in the simplified logic subgraph, and transforms the Euclidean distance into a penalty logits for the logical deviation of the candidate vocabulary through a dynamic penalty function. The penalty logits are then linearly weighted and fused with the original generation probability logits to generate the final output probability distribution subject to soft constraints of the topological structure.

[0087] The adaptive decoding penalty model employs a consistency calibration process based on reinforcement learning for parameter optimization. This process defines the procedure parsing task as a Markov decision process, treats the decoder as an intelligent agent, and uses the logical chain output by the graph inference engine as the external reward environment. Using a proximal policy optimization algorithm, the immediate reward value is calculated based on the topological fit between the generated parsing conclusion and the logical chain. The cross-attention weight matrix of the decoder is updated by maximizing the expected cumulative reward, thereby achieving end-to-end collaborative optimization between graph inference and text generation.

[0088] In this embodiment, the above mechanism performs vector-level fusion of natural language generation and the output of the graph neural network.

[0089] Step a. Decoding the hidden state and calculating the distance between graph nodes: Obtain the decoder in Hidden state vector at time step And the final graph node feature matrix output by the expansion direction one. Perform distance calculation.

[0090]

[0091] in, Indicates at time The generation state and the first Constraint Clause Node Features exist Euclidean distance in dimensional space; enumerate all valid constraint nodes. To ensure nothing is missed.

[0092] The traversal index for the feature vector dimension; Step b. Dynamic penalty logarithmic odds generation model (interacting with upstream): converts the distance calculation results into specific penalty values.

[0093]

[0094] in, For the first word in the word list The penalty log odds for each candidate word For adaptive control coefficients, and The weights and biases are used to linearly map the distance space to the vocabulary space. The upper limit of the penalty is controlled by the hyperbolic tangent function. This indicates the index of the mapping node corresponding to the word.

[0095] Step c. Final soft-constraint probability distribution model (interaction with upstream): fuse the original decoder output with the penalty term.

[0096]

[0097] in, For the final probability distribution, The original language model's unnormalized predicted logits are the odds. This represents the total size of the globally enumerated vocabulary.

[0098] In the construction and training of reinforcement learning (RL) models, the proximal policy optimization (PPO) algorithm is employed. The model state is defined as the generated sequence. Features of graphs Reward function It consists of two parts: one is the language fluency reward, and the other is the graph topology fit reward (i.e., whether the generated clause number strictly exists within the topology). (In the logical chain). During the training phase, the system continuously calculates the advantage function and updates it using gradient ascent by interacting with a power environment database containing millions of professional corpora for 500,000 episodes. The parameters of the attention layer in the large model are set with a decay factor of 0.99 and a parameter truncation. Set to 0.2.

[0099] When applying the model, no manual hard threshold configuration is required during the inference phase. It replaces the conventional Softmax algorithm, achieving smooth probabilistic intervention. If the vocabulary to be generated deviates significantly from the topology of the core subgraph clipped from direction one, It will rapidly increase and smoothly suppress the generation probability of the erroneous word, avoiding the collapse of the sentence caused by direct truncation.

[0100] This embodiment does not exist in isolation, but constitutes a closed-loop collaborative system of "dynamic supply-adaptive consumption".

[0101] The traditional pain point is that the fragments provided by the retrieval are noisy (supply-side problem), which makes the generative model prone to illusions even with constraints (consumer-side problem).

[0102] This solution aims to leverage dynamic weighting on the supply side. Based on the user's specific problem, it transforms a previously complex static knowledge network into a clear and highly targeted master-slave logical chain in real time. This high-purity chain characteristic ( The first direction (direction 1) is directly input as the absolute anchor point parameter to the second direction (adaptive decoding penalty). The second direction operates on the consumer side, no longer relying on simple character matching, but calculating the spatial distance between the generated vocabulary vector and the high-purity topological feature vector, and implementing precise probability penalties. The two directions interact pairwise, completely bridging the algorithmic gap between discrete structured graphs and continuous natural language generation.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent parsing of vertical category procedures based on hierarchical retrieval and graph reasoning, characterized in that, include: Extract the physical hierarchical topology of unstructured procedure documents and map text fragments to a high-dimensional feature space to generate text vector representations; Based on the physical hierarchy topology and text vector representation, implicit logical relationships containing mutually exclusive relationships between procedural clauses are identified, and directed edges with relational weight coefficients are constructed to form a nonlinear procedural knowledge graph. Based on the text vector representation, the first and second derivatives of the semantic gradient of the text features are used to dynamically lock the retrieval boundary of the candidate text block, and the interaction feature matrix between the query vector and the candidate text block within the boundary is reduced to a matching score to output the retrieval primitive. Using the retrieval primitive as the anchor point, local correlation subgraphs are extracted from the nonlinear procedural knowledge graph. A feature aggregation and discarding method is used to perform feature aggregation on low-priority node information with logically mutually exclusive correlation edges using a dynamically generated zero gating shielding coefficient. The result is a set of logical topological nodes that resolve mutual exclusion contradictions. The retrieval primitives and the set of logical topological nodes are input into a deep large language model. The low-priority node information discarded by the feature aggregation is transformed into a sparse penalty mask vector of the vocabulary dimension, and then added to the original log probability vector to force the generation probability of the illegal words to be set to zero, thus producing an analytical conclusion.

2. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The process of extracting the physical hierarchy topology of unstructured procedure documents and mapping text fragments to a high-dimensional feature space to generate text vector representations of the extracted physical hierarchy topology specifically includes: A hierarchical prediction function is established using a combination of hierarchical gated recurrent units and conditional random fields. The physical hierarchy topology is extracted by assigning corresponding weight coefficients to the layout feature functions corresponding to each layout feature and performing a summation operation.

3. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The process of identifying implicit logical relationships among procedural clauses based on the physical hierarchy topology and text vector representation, and constructing directed edges with relational weight coefficients to form a nonlinear procedural knowledge graph, specifically includes: Utilize a hard constraint recognizer based on regular expressions to capture explicit reference relationships between procedure clauses; By using a deep learning-based weak association classifier, we can identify semantically overlapping logical implicit relationships by calculating the divergence of entity distributions among different clauses.

4. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The process of dynamically locking the retrieval boundary of candidate text blocks based on the text vector representation, using the first and second derivatives of the semantic gradient of the text features, and reducing the dimensionality of the interaction feature matrix between the query vector and the candidate text blocks within the boundary to a matching score to output the retrieval primitive, specifically includes: The dot product of the query vector and the transpose of the candidate text block key feature matrix is ​​normalized using a normalized exponential function; The interaction feature matrix is ​​obtained by multiplying the normalized result by the value matrix; Perform mean pooling on the interaction feature matrix, multiply the pooling result by the linear projection weight matrix and add a bias term to output the matching score in one-dimensional scalar form.

5. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The process of extracting local correlation subgraphs from the nonlinear procedural knowledge graph using the retrieval primitive as an anchor point, and using a dynamically generated zero gating shielding coefficient to perform feature aggregation and discarding on low-priority node information with logically mutually exclusive correlation edges, outputting a set of logically topological nodes that resolve mutual exclusion contradictions, specifically includes: Perform a scalar multiplication operation between the gated shielding coefficient and the corresponding neighbor node features; Perform mean aggregation on the features after multiplication to obtain the aggregated neighbor feature vector; The aggregated neighbor feature vectors are concatenated with the feature vectors of the previous layer nodes, and the updated node features are obtained by multiplying them by a weight matrix and a nonlinear activation function in sequence.

6. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The process of inputting the retrieval primitives and the logical topology node set into a deep large language model, converting the low-priority node information discarded by feature aggregation into a vocabulary-dimensional sparse penalty mask vector, and adding it to the original log-probability vector to force the generation probability of violating words to be set to zero, and producing an analytical conclusion, specifically includes: For illegal words that are strongly related to the low-priority node information, the element value corresponding to it in the sparse penalty mask vector of the word list dimension is set to a minimum value that tends to negative infinity, so that the generation probability after addition and normalization exponential function processing approaches zero. For words that are unrelated to the low-priority node information, the element values ​​in the sparse penalty mask vector of the vocabulary dimension are kept to zero.

7. The intelligent parsing method for vertical category procedures based on hierarchical retrieval and graph reasoning according to claim 1, characterized in that, The method further includes: The deep large language model simultaneously outputs a visual derivation path corresponding to the parsing conclusion. The visual derivation path is presented in the form of a directed acyclic graph, where nodes represent procedural clauses and edges represent reasoning logic.

8. A vertical category procedure intelligent parsing system based on hierarchical retrieval and graph reasoning, used to implement the method of any one of claims 1 to 7, characterized in that, The system includes: The document structuring module is used to extract the physical hierarchical topology of unstructured procedure documents and map text fragments to a high-dimensional feature space to generate text vector representations. The knowledge graph construction module is used to identify implicit logical associations containing mutually exclusive logical associations between procedure clauses based on the physical hierarchy topology and the text vector representation, and to construct directed edges with relation weight coefficients to form a nonlinear procedure knowledge graph. The dynamic hierarchical dual-path retrieval module is used to dynamically lock the retrieval boundary based on the text vector representation, using the first and second derivatives of the semantic gradient of the text features, and reduce the dimensionality of the interaction feature matrix within the boundary to the matching score to output the retrieval primitive. The conflict resolution graph reasoning module is used to extract local correlation subgraphs in the nonlinear procedure knowledge graph using the retrieval primitives as anchors. It uses a dynamically generated zero gating shielding coefficient to perform feature aggregation and discarding on low-priority node information with logically mutually exclusive correlation edges, and outputs a set of logical topological nodes that resolve mutual exclusion contradictions. The structured intelligent parsing generation module is used to input the retrieval primitives and the logical topology node set into a deep large language model, convert the low-priority node information that is discarded by feature aggregation into a sparse penalty mask vector of the vocabulary dimension, and add it to the original log probability vector to force the generation probability of the illegal words to be set to zero, thus producing a parsing conclusion.

Images