A knowledge processing method, system, and storage medium based on multidimensional semantic enhancement

Abstract

This invention relates to the fields of artificial intelligence and data processing technology, specifically a knowledge processing method, system, and storage medium based on multidimensional semantic enhancement. The method includes acquiring a set of semantically segmented text; invoking a preset vectorization model to map the set of segmented text to a high-dimensional semantic vector space; invoking a trained pre-trained language model; and establishing a unique identifier mapping between the high-dimensional semantic vectors in the vector index library and the semantic entity nodes in the directed topological graph. This invention forms a dual-driven knowledge storage architecture by constructing a vector index library and a directed topological graph and establishing an association mapping between the two. This enables computers to recognize and utilize deep relationships such as preconditions and causality between knowledge entities, solving the problem of singular data representation in existing technologies.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a knowledge processing method, system and storage medium based on multidimensional semantic enhancement. Background Technology

[0002] Currently, artificial intelligence technologies, represented by large-scale language models, have demonstrated powerful capabilities in text generation. However, when processing unstructured data in specialized fields, existing technologies typically only vectorize the original documents and store them in a single index, failing to accurately represent the logical dependencies between knowledge entities. This results in computers understanding knowledge at the level of word similarity rather than logic. Furthermore, large-scale language models often lack effective logical verification mechanisms when generating responses, leading to outputs that contradict the facts in the source documents. Existing solutions rely solely on soft constraints through prompts, lacking deterministic algorithmic interception methods. Even when providing citation sources, existing solutions can only locate the source at the paragraph or document level, failing to achieve precise tracing at the sentence or even spatial coordinate level, making it difficult for users to quickly verify the authenticity of information. Most processing systems employ fixed similarity thresholds or unchanging processing logic, unable to dynamically adjust strategies based on the attributes of the content being processed, resulting in rigid performance when faced with content requiring varying levels of rigor.

[0003] Specifically, existing technologies typically only convert documents into vectors and store them in a database. This single representation method leads to a separation of semantics and logic; that is, computers can retrieve texts that are similar in wording, but cannot understand the inherent dependencies and structural relationships between the knowledge points carried by the text. Furthermore, existing technologies do not adequately support complex queries. For deep problems that require logical reasoning or multi-step connections, semantic similarity matching alone cannot provide accurate and complete context.

[0004] Therefore, existing technologies urgently need a multi-dimensional semantic enhancement knowledge processing method that can transform unstructured data into a structured semantic space and has the ability to verify logical consistency and accurately trace the source. Summary of the Invention

[0005] In this section, as well as in the abstract and title of this application, some simplifications or omissions may be made to avoid obscuring the purpose of this section, the abstract, and the title of this application, and such simplifications or omissions shall not be used to limit the scope of the invention.

[0006] To address the shortcomings of existing technologies, one objective of this invention is to provide a knowledge processing method based on multidimensional semantic enhancement.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a knowledge processing method based on multidimensional semantic enhancement, comprising: acquiring a set of semantically segmented text, wherein the set of segmented text is extracted from the original unstructured data; calling a preset vectorization model to map the set of segmented text to a high-dimensional semantic vector space, and constructing a vector index library; calling a trained pre-trained language model to extract semantic entity nodes from the set of segmented text, and constructing a directed topological graph based on the logical dependencies between the extracted nodes; establishing a unique identifier mapping between the high-dimensional semantic vectors in the vector index library and the semantic entity nodes in the directed topological graph, and using the unique identifier to integrate and bind the unstructured vector data and the structured graph topological data into an index, thereby forming a dual-driven knowledge storage architecture.

[0008] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the construction of a directed topological graph based on the logical dependencies between extracted nodes includes: analyzing the contextual semantic implications between semantic entity nodes; when the first node is the semantic premise or logical basis of the second node, establishing a unidirectional topological edge from the first node to the second node in the directed topological graph, and labeling the corresponding dependency weights.

[0009] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: receiving a query request initiated by a terminal user and retrieving related candidate text blocks from the vector index library; obtaining attribute information of the associated nodes corresponding to the candidate text blocks in the directed topological graph; dynamically calculating a similarity interception threshold based on the attribute information; and inputting the candidate text blocks into a large language model for response content generation only when the similarity of the candidate text blocks meets the similarity interception threshold.

[0010] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: using a natural language reasoning model to perform logical consistency judgment on the candidate text block and the response content output by the large language model; if the logical consistency judgment result is implied, then the response content is output; if the logical consistency judgment result is contradictory, then the response content is intercepted.

[0011] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: when the response content is intercepted, extracting abnormal semantic entities from the response content; generating a negative constraint instruction carrying the abnormal semantic entities; and re-invoking the large language model to generate the corrected response content until the logical consistency judgment is passed.

[0012] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: extracting physical location metadata of each block of text when semantically segmenting the original unstructured data, wherein the metadata includes page identifier and coordinate features; extracting physical location metadata corresponding to the block of text referenced by the response content when outputting response content; and highlighting the original document coordinates on the terminal interface according to the physical location metadata.

[0013] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: traversing unqualified nodes in the directed topological graph using a path search algorithm; and dynamically generating a nonlinear resource distribution sequence based on the unidirectional topological edges and their corresponding dependency weights.

[0014] As a preferred embodiment of the knowledge processing method based on multidimensional semantic enhancement described in this invention, the method further includes: establishing a multidimensional state machine for each terminal user to record the weight feature values ​​of each node in the directed topological graph; and updating the weight feature values ​​in real time according to the feedback vector during the user interaction process to reflect the node mastery status.

[0015] To address the shortcomings of existing technologies, another objective of this invention is to provide a knowledge processing system based on multidimensional semantic enhancement.

[0016] To achieve the above objectives, the present invention adopts the following technical solution: a knowledge processing system based on multidimensional semantic enhancement, employing the aforementioned knowledge processing method based on multidimensional semantic enhancement, including a data modeling module for semantically segmenting raw unstructured data to construct a segmented text set, and constructing a vector index library using a vectorization model; a graph construction module for extracting semantic entity nodes from the segmented text set, constructing a directed topological graph based on the logical dependencies between nodes, and establishing an association mapping between the vector index library and the directed topological graph; and an intelligent retrieval module for retrieving data related to the user's query request. The system identifies candidate text blocks and dynamically calculates a similarity interception threshold based on the node attributes of the directed topology graph. A logic verification module is used to perform logical consistency judgment on the response content output by the large language model using a natural language inference model, and performs interception or self-correction operations when a contradiction is detected. A source tracing rendering module extracts the physical location metadata of the text blocks referenced by the response content and renders them at the corresponding original document coordinates on the terminal interface. A scheduling optimization module maintains the user's multi-dimensional state machine and updates the node weight feature values ​​to dynamically generate resource distribution sequences in the directed topology graph using a path search algorithm.

[0017] To address the shortcomings of the prior art, a third objective of this invention is to provide a computer-readable storage medium.

[0018] To achieve the above objectives, the present invention adopts the following technical solution: a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the knowledge processing method based on multidimensional semantic enhancement as described above.

[0019] The beneficial effects of this invention are as follows: By constructing a vector index library and a directed topological graph and establishing a correlation mapping between the two, a dual-driven knowledge storage architecture is formed, which elevates the computer's understanding of unstructured data from a single dimension of "semantic similarity" to a multi-dimensional semantic space that integrates "logical dependencies." This enables the computer to recognize and utilize deep relationships such as preconditions and causality between knowledge entities, solving the problem of singular data representation in existing technologies.

[0020] By setting a dynamic similarity interception threshold, natural language reasoning logic consistency judgment, and negative constraint self-correction driven by abnormal semantic entities, a three-tiered mechanism is constructed for pre-filtering, post-verification, and closed-loop correction of generated content. Unlike existing technologies that rely solely on soft constraints of prompt words, this approach enforces the implicit relationship between the generated content and the facts of the source document at the algorithmic level, significantly reducing the probability of outputting contradictory or fictitious content.

[0021] By extracting and binding the physical location metadata of text blocks during the semantic segmentation stage, and performing reverse parsing and front-end highlighting rendering during response output, a direct mapping from generated text to the original document space coordinates is established. By building a multi-dimensional state machine for each user to record the weight feature values ​​of graph nodes, and utilizing a path search algorithm to traverse unqualified nodes in the graph and plan the distribution sequence, resource scheduling is transformed from a preset linear path into a non-linear optimization process driven by individual states and knowledge topology. This ensures that the resource sequence obtained by each user accurately matches their current knowledge mastery state, avoids redundant content redundancy, and improves the overall efficiency of data processing and resource scheduling at the system level. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram illustrating the formation process of the dual-drive knowledge storage architecture based on the multi-dimensional semantic enhancement knowledge processing method of this invention.

[0024] Figure 2This is a schematic diagram illustrating the process of constructing a directed topological graph based on the logical dependencies between extracted nodes using the knowledge processing method based on multidimensional semantic enhancement of this invention.

[0025] Figure 3 This is a schematic diagram of the query and response generation process of the knowledge processing method based on multidimensional semantic enhancement of the present invention.

[0026] Figure 4 This is a schematic diagram of the composition structure of the knowledge processing system based on multidimensional semantic enhancement according to the present invention. Detailed Implementation

[0027] To make the objectives, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0029] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0030] It should be noted that the "preset vectorized model M1" and "trained pre-trained language model M2" mentioned in the embodiments of the present invention are all computer program modules that have completed all training steps, have fixed model parameters, and are in the inference state. Their specific model architecture, training dataset, and training process do not constitute a limitation on the scope of protection of the present invention. Those skilled in the art can select any suitable trained model and integrate it into the processing flow of the present invention according to actual computing power and accuracy requirements.

[0031] Example 1

[0032] Reference Figure 1 This is the first embodiment of the present invention, which provides a knowledge processing method based on multidimensional semantic enhancement to solve the problems of singular representation and logical fragmentation of unstructured knowledge resources in computer systems. It includes the steps of data acquisition, first-dimensional representation construction, second-dimensional representation construction, and multidimensional semantic association and fusion.

[0033] Specifically, data acquisition includes acquiring a set of semantically segmented text blocks D1, which has been pre-extracted from the original unstructured data D. The original unstructured data D can be documents in formats such as PDF, WORD, and PPT. In this embodiment, semantic segmentation refers to dividing a long text document into segments with complete and independent semantics based on semantic indicators such as sentence coherence and paragraph themes, rather than mechanically segmenting by word count. The specific method of semantic segmentation can be based on a pre-trained text segmentation model (such as using SBD sentence boundary detection plus coherence scoring); it can also be based on the document's hierarchical headings and numbers (e.g., taking the content under each second-level heading as a block); and in one implementation, the block size can be set to 512 or 1024 characters, with 10% overlap between adjacent blocks.

[0034] The first-dimensional representation construction steps include calling a preset vectorization model M1 to map the segmented text set D1 into high-dimensional vectors, and constructing a vector index library M1' capable of deep semantic retrieval. In this embodiment, the vectorization model M1 is a neural network model that can convert text into fixed-length numerical vectors, and its conversion result makes semantically similar text vectors geometrically close in space. Specifically, the vectorization model M1 can be an encoder based on the BERT architecture, such as the text-embedding-ada-002 model; it can also be a model based on Sentence-BERT, such as the all-MiniLM-L6-v2 model (outputting 384-dimensional vectors); in one implementation, the text can be converted into a dense vector of 768 dimensions.

[0035] The steps for constructing the second-dimensional representation include calling the trained pre-trained language model M2, extracting semantic entity nodes from the chunked text set D1, analyzing the logical dependencies between nodes, and constructing a directed topological graph D2. This graph can formally represent the structure of knowledge using points and directed edges. In this embodiment, a semantic entity node refers to an independent knowledge point or conceptual unit represented by a node structure in a computer graph database. It is a basic component of the knowledge graph and can be a noun phrase, such as "Newton's First Law"; a procedural description, such as "the steps for drawing a concave lens imaging ray diagram"; or a question-and-answer pair, such as "What is photosynthesis?" and its answer forming a question-and-answer node. Logical dependencies refer to the objectively existing prerequisite logic between knowledge points that directly supports their comprehensibility or derivability, used to define the direction and type of navigation edges between nodes in the graph. Optional logical dependencies in this embodiment include prerequisites, inclusions, and causal derivations. "Prerequisite" means that you must master A before you can learn B (e.g., "linear algebra" is a prerequisite for "machine learning"); "included in" means that A is a component of the concept of B (e.g., "CPU" is included in "von Neumann architecture"); "causal derivation" means that theory A is the basis for deriving theorem B (e.g., "Maxwell's equations" derive "the existence of electromagnetic waves").

[0036] A directed topological graph D2 is a computer-readable graph data structure consisting of nodes representing semantic entities and arrowed edges representing logical dependencies. The arrows on the edges indicate the direction of the logical dependency (from premise to conclusion). A directed topological graph D2 can be stored in the Neo4j graph database, with nodes having name and type attributes and edges having weight attributes; it can also be stored in a NetworkX object, saved in a Python pkl file; in a visualization implementation, the graph can be presented as a hierarchical, interactive network graph with clearly defined arrow directions.

[0037] The multidimensional semantic association fusion step includes establishing an association mapping between the vector index library M1' and the directed topological graph D2, that is, linking a text vector representing the same concept with a node in the graph using a unique identifier, ultimately forming a dual-driven knowledge storage architecture S. In this embodiment, the association mapping refers to the mechanism of establishing a unique identifier (ID) link between text blocks in the vector index library M1' and nodes in the graph database. Specific implementation methods include storing this in a mapping table in the relational database, containing a composite primary key of chunk_id and node_id; alternatively, adding a metadata field named graph_node_id to each vector in the vector database; and also attaching a source_chunk_ids list to the graph node to directly store all text block IDs referencing this node.

[0038] This embodiment combines two heterogeneous databases through association mapping. During the construction phase, whenever a chunk is processed, its ID in the vector index database M1' and the semantic node ID extracted from the graph database are simultaneously retained in the system and associated. During the query phase, when data is found in either database, the system can instantly jump to the other database to retrieve the associated data. For example, mapping based on shared IDs can be used: when generating chunks and extracting nodes, the same globally unique ID is forced to be used, achieving a natural association. Mapping based on external mapping tables can also be used, i.e., maintaining an independent SQLite database table with fields (chunk_id, node_id), performing association queries during queries; furthermore, mapping based on vector metadata can be used to store a list of associated node_ids in the metadata field of the vector database.

[0039] Specifically, establishing the association mapping between the vector index library M1' and the directed topological graph D2 refers to assigning a globally unique node_id to each knowledge node in the directed topological graph D2 in the underlying database of the system (such as the linkage between Neo4j and Milvus, or a hybrid graph vector database). Simultaneously, in the vector index library M1', this node_id is stored as metadata for the corresponding high-dimensional semantic vector of the segmented text. When the system retrieves data, it can quickly locate the node_id through vector similarity and also trace its context vector backward from the node_id through the topological edge relationships of the directed topological graph D2, thereby achieving data collaboration in the dual-drive architecture S.

[0040] In one specific implementation, a scheme based on an external mapping table can be used to process a 500-page textbook of "University Physics". When processing the "Kepler's Laws" paragraph on page 128, a vector index library M1' is generated with the record id: v_chk_001. At the same time, the pre-trained language model M2 extracts the directed topological graph node D2 with the node id: n_kepler, and inserts (v_chk_001, n_kepler) into the mapping table. When a user asks "the three laws of planetary motion", the vector retrieval hits v_chk_001. The system finds the n_kepler node through the mapping table and traverses its predecessor node n_newton (Newton's laws of motion). The text blocks associated with n_newton, such as v_chk_000, are also added to the model's reasoning context, achieving knowledge and contextual synergistic enhancement.

[0041] The dual-drive knowledge storage architecture S can be understood as a hybrid storage and computing architecture. When processing queries, it can simultaneously or sequentially utilize the semantic search capabilities of the vector index library M1' and the logical association capabilities of the directed topological graph D2, and dynamically fuse the results of both. Specifically, when executing a query, the vector index library M1' first retrieves the preceding text block, then finds the corresponding graph node through the mapping table, and retrieves its preceding and following knowledge nodes along the outgoing and incoming edges. The text blocks associated with these nodes are then sent into the context, expanding the breadth and depth of the search.

[0042] How it works: First, the input and preparation phase: The system receives a PDF document, and the background chunking algorithm (such as Semantic Chunking) has already divided it into segments based on semantic integrity, namely the chunked text set D1, which is ready as input.

[0043] Next, in parallel modeling path one: the system calls a pre-defined, parameter-frozen vectorized model M1 (such as Sentence-BERT). Each text block is input into the model, and the output is a fixed-length numeric vector (such as a 768-dimensional array). All vectors are stored in a dedicated vector database, forming a vector index library M1'. This library can quickly find content with similar meanings. Simultaneously, in parallel modeling path two: the system also calls a pre-trained language model M2 (such as a fine-tuned version of GPT or BERT series models) with deep semantic understanding capabilities, inputs carefully designed prompts into this model, and requires the model to complete two tasks.

[0044] Task 1: Identify and extract core concepts from the text as semantic entity nodes. For example, extract nodes such as "acceleration," "net force," and "mass" from a text about "Newton's Second Law." Task 2: Analyze the relationships between these nodes. The model will determine from the context that "to understand 'acceleration,' one must first understand the concepts of 'velocity' and 'time,'" thus identifying a logical dependency between them. Based on this, the model will output a relational triple, such as (velocity, is a prerequisite for..., acceleration).

[0045] The second stage is graph construction: the system receives the relations output by the pre-trained language model M2, treats each semantic entity as a node, and treats the identified logical dependencies as unidirectional edges from premise to conclusion, and constructs the directed topological graph D2 of the entire document in the graph database.

[0046] Then, the system enters the association mapping stage: Through ID association, the system binds the "acceleration" node in the directed topology graph D2 to all text block vectors in the vector index library M1' that define "acceleration" or provide detailed explanations of it, ultimately forming a dual-drive architecture. At this point, a hybrid knowledge storage architecture is formed. When subsequent queries are initiated, the system can utilize both the vector index library M1' for semantic divergent searching and the directed topology graph D2 for logical guidance and contextual expansion.

[0047] In summary, this embodiment expands the computer's understanding of unstructured data from a single "semantic similarity" dimension to a multi-dimensional dimension that includes logical dependencies, making the data representation closer to the knowledge origins of the real world. Furthermore, the "dual-drive" architecture formed by association mapping lays a deterministic and structured data foundation for subsequent tasks such as intelligent retrieval, content generation, and path planning with logical reasoning capabilities, and solves concurrency problems such as model generation illusion and irrelevance caused by the lack of data structure.

[0048] Example 2

[0049] Reference Figure 2 Unlike the previous embodiment, this embodiment provides a preferred method for constructing a directed topological graph D2 based on the logical dependencies between nodes, solving the problem of how to accurately define and quantify the direction and strength of the logical dependencies between semantic entity nodes in the graph.

[0050] The specific method includes: first, analyzing the contextual semantic implication relationships between semantic entity nodes. Here, semantic implication refers to the logical relationship that the truth of one statement necessarily leads to the truth of another statement.

[0051] When the first node is the semantic premise or logical basis of the second node, a one-way topological edge is established in the directed topological graph, pointing from the first node to the second node, and the corresponding dependency weight is labeled. Here, the dependency weight is a numerical parameter that quantifies the strength of the support provided by the first node to the second node.

[0052] How it works: The system processor calls the pre-trained language model M2 to construct a "premise-hypothesis" statement pair from the context of a pair of semantic entity nodes for analysis. For example, if it is determined that "mastering the Pythagorean theorem" is a prerequisite for "learning the cosine theorem," then a one-way edge is established from the "Pythagorean theorem" node to the "cosine theorem" node. At the same time, the pre-trained language model M2 outputs a dependency weight value between 0 and 1 and labels it on the edge based on the degree of determinism of the premise to the conclusion (such as "must master" or "recommended to learn").

[0053] In summary, by analyzing semantic implications and quantifying dependency weights, the constructed directed topological graph D2 can not only express the presence or absence of relationships, but also accurately express the strength of those relationships, providing a more refined decision-making basis for subsequent graph-based data traversal and path planning.

[0054] Example 3

[0055] Reference Figure 1 and Figure 2 This embodiment, based on Embodiments 1 and 2, automatically constructs a "dual-driven" knowledge base in a professional field according to an implementation scenario, aiming to solve the problem of automatically converting unstructured textbook documents into a structured knowledge base. The scenario is assumed to be an administrator uploading a 500-page PDF textbook titled "Introduction to Artificial Intelligence".

[0056] First, semantic segmentation is performed. The processor calls the document parsing engine to convert the PDF content into a text stream. The system uses a semantic integrity-based segmentation algorithm, such as segmenting when the next paragraph or chapter title is detected, generating a set of 3000 segmented texts D1{C1, C2, ..., C3000}. Simultaneously, the physical location metadata M of each segment is extracted and recorded; for example, the metadata for C1 is {Page_Num:1, BBox:(50, 100, 500, 200)}.

[0057] Next, a vector index library M1' is constructed. 3000 text blocks are fed into a vectorization model M1 (e.g., text-embedding-ada-002), mapping each text block to a 1536-dimensional high-dimensional vector. All vectors and their corresponding text blocks are stored in a vector database, thus constructing the vector index library M1' capable of semantic similarity retrieval.

[0058] Next, a directed topological graph D2 is constructed. The processor calls the pre-trained large language model M3 and executes a graph construction task chain. First, core semantic entity nodes, such as "supervised learning," "unsupervised learning," "neural network," and "backpropagation algorithm," are extracted from all text blocks, totaling 1000 nodes, and stored in the graph database. Then, the system prompts the large language model M3 to analyze the logical dependencies between node pairs. For example, when analyzing "backpropagation algorithm" and "neural network," the model determines from the context that "neural network" is a logical premise of "backpropagation algorithm," so a unidirectional edge is established in the graph database from "neural network" to "backpropagation algorithm," and its dependency weight is calculated to be 0.9 (strong dependency).

[0059] Finally, a dual-drive system is established through a mapping: the system creates a unique identifier mapping between each text block in the vector index M1' and its corresponding node in the graph. For example, if text block C156 is the definition of "backpropagation algorithm," then its ID and the ID of the graph node "backpropagation algorithm" form a record pair in the mapping table. At this point, the dual-drive knowledge storage architecture S is complete. When a user searches, they can not only find semantically similar text in the vector library but also understand its context through the graph.

[0060] In summary, the textbook knowledge system, which originally required weeks of manual compilation, can be automatically transformed into a dual-driven knowledge storage architecture S with semantic search and logical reasoning capabilities within tens of minutes.

[0061] Example 4

[0062] Reference Figure 3 Unlike the previous embodiment, this embodiment aims to solve the "illusion" problem that arises when the large language model M3 generates content based on a private knowledge base, including steps such as similarity interception threshold, logical consistency judgment, and negative constraint self-correction.

[0063] Specifically, in the similarity interception threshold stage, the system receives query requests initiated by end users and retrieves related candidate text blocks D3 from the vector index library M1'; it obtains the attribute information of the associated nodes corresponding to candidate text blocks D3 in the directed topological graph D2; it dynamically calculates the similarity interception threshold based on the attribute information, and only inputs candidate text blocks D3 into the large language model M3 when the similarity of candidate text blocks D3 meets the similarity interception threshold. "Attribute information of associated nodes" refers to the node's type in the graph (e.g., "definition class," "theorem class," "case illustration class") and topological importance (e.g., in-degree, out-degree). "Dynamically calculating the similarity interception threshold" means that the system does not use a fixed value, but automatically adjusts the threshold according to the node attributes: for high-rigor nodes, the interception threshold is automatically increased to ensure a high degree of content matching; for low-rigor nodes, the threshold is appropriately decreased to allow a wider range of related content to enter.

[0064] The similarity interception threshold is dynamically calculated based on attribute information, specifically including: calculating the in-degree InD and out-degree OutD of the associated semantic entity nodes in the candidate text block D3 in the directed topological graph D2. The importance weight of the node is calculated as W1 = α·InD + β·OutD, where α and β are preset normalization coefficients (e.g., α = 0.01, β = 0.02), and the maximum value of W1 is truncated to 0.05. The node type is determined: if the node type belongs to a preset set of core concepts (such as core terms in legal provisions or core medical terms), the type weight W2 is assigned a first preset value of 0.15; if it belongs to a secondary explanatory node, the type weight W2 is assigned a second preset value of 0.05.

[0065] The final dynamically adjusted threshold T = T0 + W1 + W2, where T0 is the basic similarity threshold (e.g., 0.7).

[0066] Next, logical consistency is determined. A natural language inference model is used to assess the logical consistency between candidate text block D3 and the response content output by the large language model M3. If the logical consistency result is implication, the response content is output; otherwise, the response content is blocked. The natural language inference model is a neural network classifier specifically designed to determine the logical relationship between two text segments. Logical relationships between two text segments include implication, contradiction, and neutrality.

[0067] Then comes the negative constraint self-correction: when the response content is intercepted, abnormal semantic entities are extracted from the response content, i.e., terms or concepts that do not exist in the source document; a negative constraint instruction carrying the abnormal semantic entity is generated, and the large language model M3 is called again to generate the corrected response content until it passes the logical consistency judgment. Among them, the negative constraint instruction is a specially constructed prompt word used to explicitly tell the model that the use of a specific word or concept is prohibited.

[0068] How it works: After a user submits a query, the vector index library M1' retrieves several relevant text blocks. The system then finds the "Newton's First Law" node in the corresponding graph for these blocks, which is classified as "theorem" and has a high in-degree (depended on by many other concepts). Based on this, the system automatically increases the similarity threshold from the default 0.7 to 0.85, filtering out text blocks with scores below this threshold and sending high-quality segments to the large language model M3.

[0069] After the large language model M3 generates a response, the system uses the retrieved original text fragment as the "premise (P)" and the model's response as the "hypothesis (H)," and sends them to the Natural Language Inference (NLI) model for calculation. If the NLI model determines a "contradiction," such as the original text stating "A leads to B," but the model generates "A does not lead to B," then the system triggers an interception.

[0070] Step 3: The system extracts "abnormal semantic entities" that are not present in the original text from the intercepted answer, and generates an instruction: "Your answer just now used the concept of 'relativistic effects,' which is not mentioned in this textbook. Please restate it using only the provided reference passage." Then, the large language model M3 is called again to generate the response, and the system is validated again until it passes.

[0071] In summary, by constructing a triple closed-loop defense mechanism—pre-emptive dynamic gating filtering, post-emptive logical consistency verification, and abnormal entity-driven self-correction—the rigor of the generated content and its fidelity to the source document are enforced at the algorithmic level, significantly suppressing the illusion phenomenon of large models.

[0072] Example 5

[0073] This embodiment aims to solve the problems of coarse granularity of content citation tracing and inability to perform precise spatial positioning in the prior art, including the steps of metadata extraction and recording, output tracing and rendering.

[0074] Specifically, metadata extraction and recording involves extracting the physical location metadata of each text block during semantic segmentation of the original unstructured data D. This includes page identifiers and coordinate features. Physical location metadata refers to the precise spatial coordinate information of the text within the original PDF or document.

[0075] Among them, output tracing and rendering involves extracting the physical location metadata corresponding to the block text referenced by the response content when outputting the response content; and highlighting and rendering the corresponding original document coordinates on the terminal interface based on the metadata.

[0076] How it works: During the document parsing and chunking stage, the system generates an enhanced index for each text chunk, for example, recording that the chunk is located at "page 5, coordinates (100, 200, 400, 250)". When the model generates an answer and includes citation tags, the backend parses the tags and retrieves the corresponding coordinates from the vector library. The frontend rendering engine then draws a semi-transparent highlight box at coordinates (100, 200) on page 5 in the document viewer, allowing the user to instantly see the original source of the answer.

[0077] In summary, by extracting and recording metadata and outputting traceability and rendering, a qualitative leap has been achieved from "paragraph-level citation" to "coordinate-level traceability," making every piece of evidence in the generated content traceable and verifiable, which greatly improves the efficiency of information verification and the credibility of the system in professional scenarios.

[0078] Example 6

[0079] Reference Figures 1-3 This embodiment is based on Embodiments 4 and 5, and aims to address the accuracy and traceability issues in intelligent question answering. The scenario is as follows: a client user asks, "Why is the backpropagation algorithm effective?"

[0080] First, the system converts the question into a query vector and retrieves five candidate text blocks D3{CC1...CC5} from the vector database. Through association mapping, the system finds that these blocks correspond to the "backpropagation algorithm" node in the graph, which has the attribute of "core theorem class" and high topological importance. The system uses a dynamic threshold function T=0.7+0.15 (type weight)+0.05 (importance weight)=0.9. Among them, CC5 has a cosine similarity of only 0.85 with the question, which is lower than the interception threshold of 0.9, and is directly filtered out. Only CC1~CC4 are sent to the large language model M3.

[0081] Secondly, the large language model M3 generates an initial answer: "The backpropagation algorithm, proposed by Hinton, effectively calculates the weight gradients of each layer through the chain rule of partial derivatives." The system then initiates NLI posterior verification. CC1-CC4 are used as "premise P," and the generated answer is used as "hypothesis H," both fed into the NLI model. The NLI model outputs: implication probability is 0.97, and contradiction probability is 0.01. The judgment result is "implication," indicating logical consistency.

[0082] Next, abnormal entity detection and self-correction are performed: (Assuming in another interaction) When the large language model M3 generates an answer, it automatically adds "and the cooperation of momentum optimizer," but "momentum optimizer" is not a concept in the source document. The system detects "momentum optimizer" as an abnormal entity during entity alignment verification and immediately blocks the answer. The system automatically generates a negative constraint instruction: "You have generated an abnormal entity 'momentum optimizer,' which is not in the reference material. Please remove it and rewrite based solely on the material." Then, the model is called again, and the newly generated answer passes the verification.

[0083] Finally, the verified answer is sent to the backend. The system parses the text block ID referenced in the answer and extracts its physical location metadata from the database, such as {Page_Num:128, BBox:(55, 210, 450, 260)}. The front-end interface automatically highlights a rectangular area from (55, 210) to (450, 260) in yellow on page 128 of the PDF "Introduction to Artificial Intelligence" viewed by the user, enabling real-time and accurate traceability of the answer's basis.

[0084] In summary, dynamic gating and NLI logic collision ensure the logical consistency and factual accuracy of the answer; coordinate-level tracing makes the answer credible, verifiable, and traceable.

[0085] Example 7

[0086] This embodiment aims to solve the problem that existing resource distribution systems have fixed strategies and cannot optimize resource scheduling based on the dynamic status of individual users. It includes steps such as constructing a multi-dimensional state machine, updating state weights in real time, and generating non-linear resource distribution sequences.

[0087] First, a multidimensional state machine is constructed for each end user to record the weight feature values ​​of each node in the directed topological graph D2. A "multidimensional state machine" is a computational model that records multiple state variables and can transition between these states based on input events; here, it is used to represent the user's knowledge mastery profile.

[0088] Secondly, the state weights are updated in real time. The weight feature values ​​are updated in real time based on the feedback vector during the user interaction process to reflect the node's mastery status. The feedback vector is a parameter group composed of multi-dimensional data such as the user's answer accuracy, interaction time, and self-assessment confidence value.

[0089] Finally, a non-linear resource distribution sequence is generated. A path search algorithm is used to traverse unqualified nodes in the directed topological graph D2. Based on the unidirectional topological edges and their corresponding dependency weights, a non-linear resource distribution sequence is dynamically generated. The "non-linear resource distribution sequence" refers to a personalized learning resource recommendation list that is not ordered by chapter, but rather dynamically planned based on graph dependencies and user weaknesses.

[0090] How it works: A state machine is created for each user, where each knowledge point node in the knowledge graph has a mastery weight ranging from 0 to 1. If a user answers incorrectly multiple times in the test for the "Law of Sines" node, the feedback vector causes the weight of that node to drop to 0.4. This triggers the path planning algorithm, which scans the knowledge graph and finds that the "Law of Sines" is a prerequisite for many subsequent nodes and has a current weight below the passing threshold of 0.6. Based on this, the algorithm generates a resource distribution sequence: prioritizing the distribution of basic concepts and targeted exercises for the "Law of Sines," and suggesting reviewing its prerequisite node, "Triangle Ratio." This sequence is dynamically generated and entirely personalized.

[0091] In summary, by using state machine mechanisms and graph traversal algorithms, the system achieves fully dynamic and personalized resource distribution. Like a personal navigator, the system can automatically identify each user's weak points and plan the most efficient knowledge replenishment route, significantly improving resource utilization efficiency and scheduling intelligence.

[0092] Example 8

[0093] This embodiment is based on Embodiment 7, which aims to solve the path planning problem for personalized distribution of learning resources. The scenario assumes that user ZhangSan's mastery of mathematical knowledge is severely unbalanced.

[0094] First, the system establishes a multidimensional state machine for user ZhangSan, whose state space consists of the weight feature values ​​of all nodes in the knowledge graph. The initial states are as follows: "Definition of Trigonometric Functions" node weight 0.9, "Law of Sines" 0.8, "Law of Cosines" 0.3, "Application of Solving Triangles" 0.1.

[0095] Secondly, the user submitted consecutive incorrect answers in an exercise on "Applications of Solving Triangles," and the system collected a feedback vector {accuracy: 20%, time taken: 180 seconds, self-assessment confidence: low}. The weight of the user's "Applications of Solving Triangles" node was further reduced from 0.1 to 0.05. Simultaneously, because this error pattern triggered association analysis, the weight of its preceding node, "Law of Cosines," was also adjusted from 0.3 to 0.25.

[0096] Next, the system's path search agent detected several nodes with weights below the threshold of 0.6. By traversing the directed topology graph D2, the agent discovered that to master the "Application of Solving Triangles," one must first master the "Law of Cosines" (weight 0.9) and then the "Law of Sines" (weight 0.8). Based on this, the agent generated a completely non-linear resource distribution sequence: Step 1: Distribute the "Quick Review" material for the "Law of Sines" (its weight is 0.8, slightly below the threshold, requiring only reinforcement); Step 2: Distribute the "Key Explanation" video and "Targeted Training Problems" for the "Law of Cosines" (its weight is 0.25, a prerequisite for weak core concepts). Step 3: After the first two steps are completed and the node weights meet the requirements, distribute the comprehensive exercises on "Applications of Solving Triangles". This sequence completely skips the "Definition of Trigonometric Functions" chapter, which has already been thoroughly mastered.

[0097] In summary, by accurately characterizing user profiles through state machines and driving resource scheduling with graph path search algorithms, non-linear resource distribution with a "one person, one policy" approach is achieved, maximizing resource push efficiency and avoiding the consumption of invalid or duplicate content.

[0098] Example 9

[0099] Reference Figures 1-4 This embodiment, based on Embodiments 1 to 7, provides a knowledge processing system based on multi-dimensional semantic enhancement. This system aims to transform the aforementioned technical solutions into a physical device deployable and operable in a production environment through the organic synergy of six functional modules, achieving fully automated processing across the entire chain, from unstructured data to structured knowledge representation, from intelligent retrieval to logical verification, and from precise tracing to dynamic distribution. Specifically, the system includes a data modeling module 100, a graph construction module 200, an intelligent retrieval module 300, a logical verification module 400, a tracing and rendering module 500, and a scheduling optimization module 600.

[0100] Specifically, the data modeling module 100 is configured to: receive the uploaded raw unstructured data D, use a semantic integrity-based segmentation algorithm (such as scoring based on paragraph boundaries, chapter titles, or sentence coherence) to segment the long document into text blocks with independent and complete semantics, and construct a set of text blocks D1; ​​during the segmentation process, synchronously extract the physical location metadata of each text block in the original document, including page identifiers (such as page numbers) and coordinate features (such as bounding box coordinates BBox(x0, y0, x1, y1)), and bind and store this metadata as an additional field bound to the text block; call a preset vectorization model M1 (such as an encoder based on the BERT architecture) to map each text block into a fixed-dimensional high-dimensional semantic vector, and store the vector, the original text, and the metadata together in a vector database to construct a vector index library M1' that supports efficient semantic retrieval.

[0101] The data modeling module 100 provides high-quality structured input for the entire system. Semantic segmentation ensures that the contextual integrity of knowledge is not lost due to segmentation; the binding of physical location metadata lays the data foundation for subsequent accurate tracing. At the same time, the construction of the vector index library M1' enables the system to complete large-scale document semantic retrieval in milliseconds.

[0102] The graph construction module 200 is configured to: invoke a pre-trained language model M2 to perform semantic analysis on the chunked text set D1 generated by the data modeling module 100, extracting core knowledge points or concepts as semantic entity nodes. A semantic entity node is a structured data unit containing attributes such as name and type. Based on the extracted semantic entity nodes, the same or another pre-trained language model M2 is invoked to analyze the contextual semantic implications between node pairs, identifying and labeling logical dependencies between nodes. Logical dependencies include, but are not limited to, types such as "preceded by," "included in," and "causally derived." The semantic entity nodes are stored in the graph database as node entities, and the identified logical dependencies are used as unidirectional topological edges pointing from premise to conclusion. The corresponding dependency weights (a value between 0 and 1) are labeled according to the dependency strength, constructing a directed topological graph D2. An association mapping is established between the chunked text in the vector index library M1' and the semantic entity nodes in the directed topological graph D2. In practical implementation, a bidirectional traceable link can be formed between the two by adding graph node IDs to the vector metadata and recording the corresponding relationships in the maintenance mapping table.

[0103] The graph construction module 200 constructs a second-dimensional knowledge representation, namely a logical topological graph, in addition to the vector index. Through association mapping, the system integrates semantic similarity retrieval and logical dependency reasoning into a whole, forming a "dual-drive knowledge storage architecture S", which enables the computer's understanding of knowledge to leap from a single "similarity" to "similar and related".

[0104] The intelligent retrieval module 300 is configured to: receive query requests initiated by end users; call a preset vectorization model M1 to convert the query text into query vectors; perform an approximate nearest neighbor search in the vector index library M1' to retrieve a set of related candidate text blocks D3; obtain the attribute information of the associated nodes corresponding to each candidate text block D3 in the directed topological graph D2 through the association mapping established by the graph construction module 200. The attribute information includes node type (such as "definition class", "theorem class", "procedural rule class") and / or topological importance indicators (such as the node's in-degree, out-degree, and betweenness centrality); based on the obtained node attribute information, call a preset dynamic threshold function to calculate the similarity interception threshold applicable to the current query in real time. The input of this function is the node attribute features, and the output is the dynamically adjusted threshold; only candidate text blocks D3 whose similarity meets the dynamic similarity interception threshold are admitted and passed to the downstream large language model M3 as the generation context.

[0105] The intelligent retrieval module 300 realizes intelligent and dynamic retrieval strategies. By allowing the structural information of the knowledge graph to guide the admission criteria for semantic retrieval, the system can adaptively adjust the screening strictness when faced with knowledge content of different types and importance, thereby ensuring that the context input to the generation model achieves the optimal balance in relevance, accuracy, and purity.

[0106] The logic verification module 400 is configured to: acquire the response content output by the large language model M3, and construct a Natural Language Inference (NLI) input pair with the candidate text block D3 content admitted by the intelligent retrieval module 300. Here, the candidate text block D3 serves as the "premise," and the response content as the "hypothesis." The trained NLI model is invoked to classify the logical relationship between the input pair, determining whether there is an implication, contradiction, or neutral relationship between the premise and the hypothesis. If the judgment result is "implication," it indicates that the response content can be logically derived from the source document, and the response content is allowed to be output; if the judgment result is "contradiction," it indicates that there is a logical conflict between the response content and the source document, and an interception operation is triggered.

[0107] Optionally, if the judgment result is "neutral", the system can add a prompt mark and allow the response, or perform a more stringent secondary verification. When the response content is intercepted, the system further extracts abnormal semantic entities (i.e., terms or concepts that do not exist in candidate text block D3) from the response content, automatically generates negative constraint instructions carrying the abnormal entities, calls the large language model M3 again and requires it to remove the abnormal entities and regenerate, until the generated response content passes the logical consistency judgment or reaches the preset maximum number of retries.

[0108] The logic verification module 400 constructs a "post-audit" mechanism for the generated content. By introducing an NLI model for objective logical comparison, it elevates illusion detection from an uncertain "soft cue" to a quantifiable "hard judgment," forcibly ensuring the logical consistency and factual fidelity of the output at the algorithmic level. The self-correction closed loop further endows the system with self-healing capabilities, preventing service termination due to a single generation failure.

[0109] The source tracing rendering module 500 is configured to: after the logic verification module 400 allows the response content, parse the segmented text identifiers referenced in the response content. These identifiers can be embedded during generation by the large language model M3 or matched retrospectively by the system; based on the extracted segmented text identifiers, query the corresponding physical location metadata from the metadata stored in the data modeling module 100 to obtain the page identifier and coordinate features; and send the physical location metadata along with the response content to the terminal interface. The terminal interface locates the corresponding page number in the original document based on the page identifier and highlights the rectangular area defined by the coordinate features (e.g., drawing a semi-transparent yellow rectangle), achieving a spatial coordinate-level visual link between the generated content and the source document evidence.

[0110] The source tracing rendering module 500 refines the granularity of source tracing from the coarse-grained "paragraph level" to the "coordinate level." As users see the response, the document automatically locates and highlights the original source of the answer, achieving a "what you see is what you get" evidence presentation. This not only improves the efficiency of information verification but, more importantly, significantly enhances user trust in the system's output through a transparent source tracing mechanism.

[0111] The scheduling optimization module 600 is configured to: establish and maintain a multi-dimensional state machine for each terminal user. The state space of this state machine is defined by the weight feature values ​​corresponding to each semantic entity node in the directed topological graph D2. The weight feature value represents the user's mastery of the knowledge point represented by the node and is a continuous numerical value (such as a floating-point number between 0 and 1). It continuously receives feedback vectors generated during user interaction (which may include multi-dimensional data such as answer accuracy, interaction duration, interaction frequency, and self-assessment confidence value), and updates the weight feature values ​​of the corresponding nodes in real time according to a preset state transition function. It uses a path search algorithm (such as breadth-first traversal or topological sorting algorithm as shown in the figure) to traverse the directed topological graph D2, detecting nodes whose weight feature values ​​are below a preset threshold. Based on the detected nodes, and combined with the unidirectional topological edges and dependency weights in the graph, it dynamically generates a resource distribution sequence. This sequence prioritizes pushing resources corresponding to key weak nodes with unmet dependencies or high dependency weights, forming a non-linear, personalized resource push list.

[0112] The scheduling optimization module 600 realizes the transformation from "passively waiting for resources to be acquired" to "actively distributing resources on demand". By meticulously depicting the knowledge mastery of each user through a state machine and allowing the graph topology structure to drive the automatic optimization of the distribution path, the system can plan the most efficient resource acquisition sequence for each user, avoiding the distribution of invalid or duplicate content and improving the efficiency of resource utilization and the level of intelligent scheduling.

[0113] Working principle: The working principle of this system can be divided into two main stages: the offline knowledge base construction stage and the online query and reasoning service stage. The former completes the transformation from raw unstructured data D to a dual-driven knowledge storage architecture, while the latter provides users with intelligent question-answering services with logical verification and accurate source tracing based on this architecture, and continuously updates user status.

[0114] Phase 1: Offline knowledge base construction.

[0115] Administrators upload raw, unstructured data D, such as a set of PDF documents containing hundreds of pages, through the management interface.

[0116] The data modeling module 100 first calls the document parsing engine to extract the text content stream of the document. Then, the module executes a semantic chunking algorithm. This algorithm does not mechanically cut the text according to a fixed number of characters, but rather performs boundary detection based on semantic integrity. Specifically, the algorithm detects the document's natural paragraph separators, chapter title markers, and semantic coherence scores between sentences, and segments the text at locations where there are significant semantic shifts or terminations, generating several chunks of text with independent and complete semantics, forming a chunk set = {d1, d2, ..., dn}.

[0117] During the text segmentation process, the data modeling module 100 extracts and binds physical location metadata for each segment. This metadata precisely records the spatial location of the text segment within the original document, including page identifiers (such as page numbers) and coordinate features (such as the coordinates of the top-left and bottom-right corners of a rectangular bounding box). For example, for a text segment extracted from page 47, its metadata is M={Page_ID:47, BBox:(x0, y0, x1, y1)}. This metadata is stored as an additional field along with the segment text itself, providing a data foundation for coordinate-level positioning by the subsequent source tracing rendering module 500.

[0118] Finally, the data modeling module 100 calls the preset vectorization model M1 to map each text block in the segmented text set into a high-dimensional semantic vector. Vectorization model M1 is a pre-trained neural network encoder with fixed parameters, capable of converting variable-length natural language text into dense numerical vectors of fixed dimensions, ensuring that semantically similar texts are geometrically close in vector space. All generated vectors, along with their corresponding original text blocks and metadata, are stored in a vector database, constructing a vector index library M1'. This index library supports fast approximate nearest neighbor search and serves as the semantic matching engine for subsequent intelligent retrieval.

[0119] At the same time or after the data modeling module 100 completes the vectorization process, the graph construction module 200 initiates the knowledge structuring process.

[0120] First, the graph construction module 200 calls the pre-trained language model M2 to perform deep semantic analysis on the segmented text set. The pre-trained language model M2 is a neural network model with powerful language understanding capabilities, and in this system it is configured to perform two association tasks.

[0121] Task 1: Semantic Entity Node Extraction. The pre-trained language model M2 identifies and extracts core concepts, terms, rules, and theorems from text, formalizing them as semantic entity nodes. Each node is a structured data unit, containing at least the node name and node type (e.g., definition class, theorem class, procedural rule class, case illustration class). For example, nodes such as "shareholder derivative suit," "preliminary procedure," and "lawsuit against invalid board resolution" are extracted from relevant texts of the Civil Code.

[0122] Task 2: Logical Dependency Identification. The pre-trained language model M2 further analyzes the contextual semantic implications between the extracted nodes. When one node constitutes a semantic premise or logical basis for understanding another node, the model identifies a logical dependency between them and determines the type and strength of the dependency. Dependency types include, but are not limited to: prerequisite (A is necessary prior knowledge for learning B), contained in (A is a component of B), causal derivation (B can be deduced from A), etc. Dependency strength is represented by a continuous numerical value between 0 and 1, called the dependency weight; the larger the value, the tighter the dependency.

[0123] Subsequently, the graph construction module 200 persistently stores the above analysis results in the graph database, constructing a directed topological graph D2. In the graph database, each semantic entity node is created as a vertex entity; each logical dependency is created as a unidirectional topological edge from the premise node to the conclusion node, with the dependency weight corresponding to the dependency marked on the edge.

[0124] After the graph construction module 200 completes the graph construction, it performs a key fusion operation: establishing an association mapping between the vector index library M1' and the directed topological graph D2.

[0125] The specific implementation of this association mapping is as follows: For each semantic entity node in the directed topological graph D2, identify the corresponding text block (i.e., the text block describing or defining the knowledge point) in the data modeling module 100, and establish a bidirectional link between their unique identifiers. In engineering implementation, the ID of the association graph node can be added to the metadata field of the vector index library M1', and the list of IDs of the association text block can be stored in the attributes of the graph node; or an independent mapping table can be maintained.

[0126] Thus, the vector index library M1' and the directed topological graph D2 are integrated through association mapping, forming a dual-drive knowledge storage architecture S. This architecture possesses dual capabilities: on the one hand, it can perform efficient semantic search through vector similarity; on the other hand, it can perform logical reasoning and path navigation through the graph topology. The association mapping between the two ensures that when either capability is triggered, the data of the other capability can be instantly accessed, achieving synergistic enhancement.

[0127] Phase Two: Online Query and Reasoning Service.

[0128] Once the offline construction phase is complete and the dual-drive knowledge storage architecture S is ready, the system enters online service mode, ready to respond to end-user query requests at any time.

[0129] End users can initiate a natural language query request Q through the client interface, such as: "If a company's board resolution violates the articles of association, can shareholders directly file a lawsuit?"

[0130] After receiving the query request Q, the intelligent retrieval module 300 performs the following steps: Step 1: Query Vectorization and Preliminary Retrieval. The intelligent retrieval module 300 calls the same preset vectorization model M1 as the data modeling module 100 to map the query text Q to the query vector Vq. Subsequently, an approximate nearest neighbor search is performed in the vector index library M1' to calculate the cosine similarity between Vq and each vector in the library, and to recall the Top-K candidate text blocks D3{C1, C2, ..., Ck} with the highest similarity and their similarity scores.

[0131] Step 2: Obtaining Graph Node Attributes. The intelligent retrieval module 300 queries the attribute information of the associated nodes in the directed topological graph D2 for each candidate text block D3 through association mapping. The attribute information includes the node type (such as "procedural rule class") and topological importance indicators (such as the centrality measures of the node in the graph, such as in-degree and out-degree).

[0132] Step 3: Dynamic Threshold Calculation. The intelligent retrieval module 300 calls a preset dynamic threshold function. This function takes the attribute information of the associated nodes of candidate text block D3 as input variables and outputs a dynamically adjusted similarity interception threshold T. The design logic of this function is as follows: for nodes with high rigor and high topological importance, the interception threshold is increased to ensure a high degree of contextual matching before entering the generation stage; for nodes with low rigor and auxiliary explanatory nature, the threshold is appropriately lowered to allow a wider range of related content to enter.

[0133] Step 4: Gating and Filtering. The module compares the similarity score of each candidate text block D3 with the dynamic threshold T one by one. Only text blocks whose similarity score is greater than or equal to T are allowed to pass through the gating; text blocks with a similarity score below the threshold are directly blocked and will not enter the downstream generation stage. For example, if the similarity score of candidate block C5 is 0.65 and the dynamic threshold T is 0.77, then C5 will be blocked.

[0134] Step 5: Context Assembly and Model Invocation. The intelligent retrieval module 300 assembles the approved candidate text block set D3 with the user's original query request Q into a structured prompt word, and invokes the large language model M3. This large language model M3 is an independent large-scale generative language model (LLM), whose parameter scale and generation capabilities far exceed those of the aforementioned vectorized model M1 and pre-trained language model M2 used for feature extraction. Based on the provided candidate text block D3 as context, the LLM infers and integrates the query question, generating preliminary response content in natural language form.

[0135] After the large language model M3 generates the initial response content, it does not return it directly to the user. Instead, it first sends it to the logic verification module 400 for post-generation verification. The function of this module is to use algorithms to forcibly verify the logical consistency between the generated content and the source document.

[0136] The logic verification module 400 obtains the candidate text block set D3 approved by the intelligent retrieval module 300 and uses it as "premise P"; at the same time, it obtains the response content generated by LLM and uses it as "hypothesis H". Natural language inference input pairs are constructed in the form of (P, H).

[0137] The logic verification module 400 calls the trained NLI model. This model is a neural network classifier specifically designed to determine logical relationships between texts, typically based on a cross-encoder architecture and fine-tuned on natural language inference datasets such as SNLI and MNLI. The model receives input pairs (P, H) and outputs probability distributions for three mutually exclusive categories: Implication: The logic of H can be derived from P; Neutral: P neither supports nor opposes H; Contradiction: H and P have a logical conflict.

[0138] The logic verification module 400 presets a contradiction probability threshold τ (e.g., 0.8). When the contradiction probability output by the NLI model is higher than τ, it is determined that the response content has a serious logical conflict with the source document, triggering an interception operation. When the implication probability is the highest and the contradiction probability is lower than τ, it is determined to be "implication", the response content passes the verification, and is allowed to proceed.

[0139] If the response content is intercepted, the logic verification module 400 initiates a self-correction process: First, a named entity recognition algorithm is invoked to extract entity sets from both the candidate text block D3 set and the response content. Through set difference operations, terms or concepts present in the response content but not in the candidate text block D3 are identified and marked as "abnormal semantic entities."

[0140] Secondly, an automatic negative constraint instruction is generated, which explicitly lists the abnormal semantic entities and requires the LLM to strictly remove these entities during regeneration, and rewrite only based on the provided candidate text block D3.

[0141] Then, the negative constraint instruction, along with the original query and candidate text block D3, is input again into the large language model M3 to obtain the corrected response content.

[0142] Finally, the response content is resubmitted to the NLI model for verification. This process is repeated until the generated response content passes the logical consistency check, or the preset maximum number of retries (e.g., 3 times) is reached, at which point a service degradation prompt is returned.

[0143] Once the response content passes the verification of the logic verification module 400 and is allowed to proceed, the source tracing rendering module 500 intervenes to add spatial coordinate-level source tracing information to the response content.

[0144] The source-tracing rendering module 500 parses the chunked text identifiers referenced in the response content. These identifiers can be automatically embedded by the large language model M3 during generation based on the underlying context (e.g., in the form of [Source_ID]), or the system can determine the correspondence between the response content and the candidate text chunk D3 a posteriori through a semantic matching algorithm.

[0145] Based on the parsed text block identifiers, the corresponding physical location metadata is queried from the metadata field of the vector index library M1' constructed by the data modeling module 100. This metadata includes the page identifier and precise coordinate features of the text block in the original document.

[0146] The physical location metadata is encapsulated along with the response content and sent to the client. Upon receiving the response content and metadata, the client's rendering engine (such as a PDF.js-based document viewer) automatically performs the following operations: locates the document to the corresponding page number based on the page identifier; overlays a semi-transparent colored highlight layer (such as a yellow rectangle with 30% transparency) on the rectangular area defined by the coordinate features; and optionally displays floating tooltips next to the highlighted area. Thus, when reading the AI-generated answer, the user can instantly see the precise location of the original text upon which the answer is based within the document, achieving spatial mapping and visualization of the "answer-evidence" relationship.

[0147] After completing a query service, the scheduling optimization module 600 asynchronously updates the user status and plans the subsequent resource distribution sequence in the background, without blocking the response speed of the current interaction.

[0148] The scheduling optimization module 600 maintains a multi-dimensional state machine for each terminal user. The state space of the state machine consists of the weight feature values ​​corresponding to each semantic entity node in the directed topological graph D2. Each weight feature value is a continuous value between 0 and 1, representing the user's mastery of the knowledge point represented by the node (0 indicates no mastery, 1 indicates complete mastery).

[0149] During user interaction, the system continuously collects feedback vectors. A feedback vector is a multi-dimensional data set that may include behavioral data such as the user's query topic, answer accuracy, interaction duration, whether they actively clicked the source link, and whether they asked follow-up questions. The scheduling optimization module 600, based on a preset state transition function, converts the feedback vectors into increments or decrements of weighted feature values, updating the weighted feature values ​​of the corresponding nodes and their associated nodes in real time. For example, if a user asks in-depth questions about "shareholder derivative suits" and clicks the source link, the weight of that node increases from 0.3 to 0.55; while the weight of its preceding node, "determination of the validity of board resolutions," decreases from 0.4 to 0.35 because the user indirectly revealed a misunderstanding in related questions.

[0150] The path search algorithm (such as breadth-first traversal, depth-first traversal, or topology sorting-based traversal algorithm as shown in the figure) is used to scan the constructed directed topological graph D2, detect all nodes with weight feature values ​​lower than the preset threshold (such as 0.6), and mark these nodes as non-compliant nodes.

[0151] Based on the detected set of non-compliant nodes, and combined with the unidirectional topological edges and their dependency weights in the directed topological graph D2, a path planning algorithm is executed. The core logic of the algorithm is to prioritize processing the weakest nodes with an in-degree of zero or the highest dependency weight. The algorithm outputs an ordered resource distribution sequence, where each item corresponds to a resource package to be pushed (such as instructional videos, practice question banks, case studies, etc.). This sequence is non-linear; that is, it does not follow the chapter order of the original document, but rather dynamically generates an optimized resource acquisition path based on the user's current weakness graph topology. For example, for user ZhangSan, the sequence is: Step 1: Send out the "Key Explanation" and "Targeted Exercises" on "Determination of the Validity of Board Resolutions".

[0152] Step 2: Once the node's weight meets the requirements, push out the "Comprehensive Case Analysis" for "Shareholder Derivative Litigation".

[0153] Step 3: Push out extended reading on "Exceptions to Shareholder Derivative Litigation".

[0154] This sequence completely skips nodes that the user already knows and nodes that have no direct topological connection to the current weak point.

[0155] The generated resource distribution sequence is pushed to the user's client, presented to the user as a personalized suggestion for subsequent learning paths or an automated resource push list.

[0156] In summary, this system completes a full closed loop, from the raw data vectorization of the data modeling module 100, the knowledge structuring of the graph construction module 200, the dynamic gating retrieval of the intelligent retrieval module 300, the NLI posterior verification and self-correction of the logic verification module 400, the coordinate-level precise source tracing of the source tracing rendering module 500, to the user state update and dynamic resource distribution of the scheduling optimization module 600. These six modules work collaboratively in both the offline construction and online inference stages, jointly achieving multi-dimensional semantic enhancement processing of unstructured knowledge.

[0157] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the knowledge processing method based on multidimensional semantic enhancement as proposed in the above embodiments.

[0158] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0159] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0160] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0161] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge processing method based on multidimensional semantic enhancement, characterized in that: include, Obtain a semantically segmented set of text blocks (D1), which is extracted from the original unstructured data (D); The preset vectorization model (M1) is invoked to map the segmented text set (D1) to a high-dimensional semantic vector space, and a vector index library (M1') is constructed. The trained pre-trained language model (M2) is invoked to extract semantic entity nodes from the segmented text set (D1), and a directed topological graph (D2) is constructed based on the logical dependencies between the extracted nodes. A unique identifier mapping is established between the high-dimensional semantic vectors in the vector index library (M1') and the semantic entity nodes in the directed topological graph (D2). The unstructured vector data and the structured graph topological data are integrated and indexed and bound through the unique identifier to form a dual-drive knowledge storage architecture (S).

2. The knowledge processing method based on multidimensional semantic enhancement as described in claim 1, characterized in that: The construction of the directed topological graph (D2) based on the logical dependencies between the extracted nodes includes: Analyze the contextual semantic implications between semantic entity nodes; When the first node is the semantic premise or logical basis of the second node, a unidirectional topological edge from the first node to the second node is established in the directed topological graph (D2), and the corresponding dependency weight is marked.

3. The knowledge processing method based on multidimensional semantic enhancement as described in claim 1 or 2, characterized in that: It also includes, Receive a query request initiated by the end user and retrieve the associated candidate text block (D3) in the vector index (M1'). Obtain the attribute information of the associated node corresponding to the candidate text block (D3) in the directed topological graph (D2); The similarity interception threshold is dynamically calculated based on the attribute information, and the candidate text block (D3) is input into the large language model (M3) to generate response content only when the similarity of the candidate text block (D3) meets the similarity interception threshold.

4. The knowledge processing method based on multidimensional semantic enhancement as described in claim 3, characterized in that: Also includes: The logical consistency between the candidate text block (D3) and the response content output by the large language model (M3) is determined using a natural language inference model. If the logical consistency determination result is implied, then the response content is output; If the logical consistency judgment result is contradictory, then the response content is intercepted.

5. The knowledge processing method based on multidimensional semantic enhancement as described in claim 4, characterized in that: Also includes: When the response content is intercepted, the abnormal semantic entities in the response content are extracted; Generate a negative constraint instruction carrying the abnormal semantic entity, and re-invoke the large language model (M3) to generate the corrected response content until the logical consistency judgment is passed.

6. The knowledge processing method based on multidimensional semantic enhancement as described in claim 4, characterized in that: Also includes: When semantically segmenting the raw unstructured data (D), the physical location metadata of each segment text is extracted, and the metadata includes page identifiers and coordinate features; When outputting the response content, extract the physical location metadata corresponding to the block text referenced by the response content; The physical location metadata is used to highlight and render the corresponding original document coordinates on the terminal interface.

7. The knowledge processing method based on multidimensional semantic enhancement as described in claim 2, characterized in that: Also includes: The path search algorithm is used to traverse the unqualified nodes in the directed topological graph (D2); Based on the unidirectional topological edges and their corresponding dependency weights, a non-linear resource distribution sequence is dynamically generated.

8. The knowledge processing method based on multidimensional semantic enhancement as described in any one of claims 1, 2, 4, 5 or 7, characterized in that: Also includes: A multidimensional state machine is established for each end user to record the weight feature values ​​of each node in the directed topological graph (D2). The weighted feature values ​​are updated in real time based on the feedback vectors during user interaction to reflect the node's mastery status.

9. A knowledge processing system based on multidimensional semantic enhancement, employing the knowledge processing method based on multidimensional semantic enhancement as described in any one of claims 1 to 8, characterized in that: include, The data modeling module (100) is used to perform semantic chunking on the raw unstructured data (D) to construct a chunked text set (D1) and to construct a vector index library (M1') using a vectorized model (M1). The graph construction module (200) is used to extract semantic entity nodes from the block text set (D1), construct a directed topological graph (D2) based on the logical dependencies between nodes, and establish an association mapping between the vector index library (M1') and the directed topological graph (D2). The intelligent retrieval module (300) is used to retrieve candidate text blocks (D3) associated with the user's query request and dynamically calculate the similarity interception threshold based on the node attributes of the directed topological graph (D2). The logic verification module (400) is used to perform logical consistency judgment on the response content output by the large language model (M3) using the natural language reasoning model, and to perform interception or self-correction operations when the judgment is contradictory; The source-tracing rendering module (500) is used to extract the physical location metadata of the block text referenced by the response content and to render the marker at the original document coordinates corresponding to the terminal interface. The scheduling optimization module (600) is used to maintain the user's multidimensional state machine and update the node weight feature values ​​to dynamically generate resource distribution sequences in the directed topology graph (D2) using a path search algorithm.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the knowledge processing method based on multidimensional semantic enhancement as described in any one of claims 1 to 8.

Images