Knowledge asset management method based on artificial intelligence multi-source business document association analysis

Abstract

The application discloses a knowledge asset management method based on artificial intelligence multi-source business document correlation analysis, and relates to the fields of knowledge management and artificial intelligence.The method unifies data of multi-source heterogeneous documents and constructs a document structure tree; generates hierarchical semantic vectors through word-level coding and structure context gate fusion; extracts entities and relations to generate knowledge triples; constructs a global knowledge graph through cross-document entity alignment, complements implicit correlations by combining a graph attention network and a rule soft priori; realizes knowledge addition, update and decay through polymorphism management according to version time sequence conflict detection; and finally provides semantic retrieval, knowledge question answering and influence analysis through graph vector double-channel fusion.The application realizes the automatic conversion of business documents into computable and traceable knowledge assets, and improves the accuracy of knowledge extraction and the intelligent level of management.

Description

Technical Field

[0001] This invention relates to the fields of knowledge management and artificial intelligence technology, and in particular to a knowledge asset management method based on artificial intelligence-based multi-source business document correlation analysis. Background Technology

[0002] As enterprises continue to increase their digitalization, they accumulate a massive amount of business documents, including contracts, bidding documents, technical solutions, meeting minutes, processes and systems, emails, project reports, and more. These documents come from diverse sources, have heterogeneous formats, complex semantic relationships, and evolve dynamically over time.

[0003] Existing knowledge management systems have the following shortcomings:

[0004] Problem 1: Isolated storage and lack of correlation. Current MIS / knowledge base systems mostly store documents by folder or tag, lacking semantic connections between documents and failing to reflect the inherent relationships within the business context. For example, the constraints between contracts and technical solutions under the same project cannot be automatically identified.

[0005] Question 2: High cost of manual tagging. The existing system relies on manual metadata tagging and classification for each document, which is inefficient, subjective, and has a slow update speed, making it difficult to cope with massive incremental data.

[0006] Question 3: The search relies primarily on keyword matching. Traditional full-text search only performs literal matching, failing to understand semantics, resulting in fragmented search results. It cannot achieve "knowledge discovery"-style associative retrieval, such as automatically finding "all documents that affect the project's schedule constraints."

[0007] Question 4: Knowledge is difficult to retain and reuse. The tacit knowledge (decision logic, rule constraints, and empirical conclusions) in the documents is not extracted in a structured way, and cannot be formed into computable and reasonable knowledge assets.

[0008] Question 5: Lack of timeliness and context adaptation. As documents evolve with business changes, the existing static knowledge base cannot automatically identify, update, and correct changes in relationships between older versions, leading to "dirty knowledge" contamination.

[0009] Therefore, there is an urgent need for a knowledge asset management method that can automatically perform semantic parsing, cross-document association, knowledge extraction, graph construction, and dynamic updates on multi-source heterogeneous business documents, so as to improve the efficiency of business knowledge utilization and the level of knowledge asset governance. Summary of the Invention

[0010] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a knowledge asset management method based on artificial intelligence multi-source business document association analysis. Through hierarchical semantic encoding with layout structure awareness, cross-document entity alignment, graph neural network association completion, and temporal evolution governance, it realizes the automatic transformation and continuous governance of business documents into structured knowledge assets, significantly improving the accuracy of knowledge extraction, the completeness of association, and the level of intelligent management.

[0012] Specifically, this invention provides the following technical solution: a knowledge asset management method based on artificial intelligence multi-source business document association analysis, comprising the following steps: Step 1: performing unified data processing on multi-source heterogeneous business documents, processing extractable text, scanned image, table-intensive, and rich media hybrid documents through text extraction, OCR recognition, table parsing, layout analysis, and OCR joint channel processing to construct a document structure tree; Step 2: based on the document structure tree, using a semantic boundary classifier to segment the document into semantic blocks carrying structural context, and fusion through word-level encoding, sentence-level weighted pooling, and structural context gating to generate hierarchical semantic vectors; Step 3: based on the hierarchical semantic vectors, extracting key business entities through a span-level named entity recognition model, and identifying relationships between entities through a prompt relationship extraction model. The semantic relationships are used to generate knowledge triples with confidence and source lineage; Step 4: Combine attribute similarity, contextual semantic similarity, and structural similarity, and combine pseudo-co-occurrence verification to normalize and align cross-document entities to construct a global knowledge graph; Step 5: Based on the entity structure embedding learned by the graph attention network, combine the translation distance model and rule soft prior to complete the relationship of the knowledge graph and generate interpretable association paths; Step 6: Detect conflicts between new and old knowledge based on document version and time sequence, execute knowledge addition, update, or decay operations through polymorphic governance state machine, and achieve soft deletion with activity weight decay to dynamically maintain the knowledge graph; Step 7: Based on the governed knowledge graph, provide semantic association retrieval and knowledge question answering through graph vector dual-channel fusion index, and implement hierarchical sharing according to lineage and permissions.

[0013] Preferably, step two, generating hierarchical-aware semantic vectors, includes: constructing a document structure tree T=(N, E), and extracting structural context vectors s for each semantic block. i Transform the plain text semantic vector t i With s i Hierarchical semantic vectors are generated through the gated fusion layer. , where the gate vector W cs W g b g is a learnable parameter, and ⊙ represents element-wise multiplication.

[0014] Preferably, the pseudo co-occurrence verification in step four includes: for candidate entity pairs whose comprehensive similarity is within the fuzzy range, calculating their co-occurrence rate Co(a, b) in multiple independent documents; when the co-occurrence rate is higher than a preset co-occurrence threshold, increasing the merging confidence of the candidate entity pair and performing merging.

[0015] Preferably, the relation completion in step five includes: modeling triples in the complex embedding space by rotation, traversing candidate entities according to their energy scoring function to complete the completion; and using business rules as soft priors, when a new triple makes the graph satisfy the rules, a reward score is given to correct the completion result.

[0016] Preferably, step five generates an interpretable association path, which includes: searching for a path connecting two entities on the knowledge graph through weighted path reasoning, the path confidence is represented by the product of the confidence of each step relationship edge and the confidence of the intermediate entity, and returning the path with the highest score and limited number of steps as the evidence chain.

[0017] Preferably, the polymorphic governance state machine in step six includes: performing addition when there is no corresponding old triplet in the graph and the comprehensive decision score is higher than the addition threshold; performing update and retaining the historical version chain when there is a conflicting old triplet and the decision score is higher than the update threshold; and setting it to a decay state when the old information is replaced or the confidence decreases, and assigning an activity weight that decays exponentially over time to achieve soft deletion.

[0018] Preferred, the comprehensive decision score is based on Calculate, where γ new For the new triple confidence level, λ d Here, dist(t) is the time decay coefficient. new , t old () represents the time difference between the old and new versions.

[0019] Preferably, step seven, the dual-channel fusion of graph vectors, includes: merging the cosine correlation of the semantic vector channel with the graph distance correlation of the graph structure channel according to... The fusion sort is used, where ρ∈[0,1] is the fusion weight.

[0020] Preferably, step seven further includes: generating an evidence chain containing source documents and page numbers from the knowledge question-and-answer results; deducing the affected downstream constraint nodes along the associated path and outputting an impact diffusion transmission diagram, triggering a risk warning when the state of key knowledge changes.

[0021] Preferably, all extraction, alignment, completion, update, attenuation, and manual review operations generated in steps one through seven are recorded as audit logs, and each knowledge edge retains the lineage information of the source document and semantic block, supporting the tracing of the original text and hierarchical sharing.

[0022] This invention provides a knowledge asset management method based on artificial intelligence-based multi-source business document correlation analysis, which has the following beneficial effects:

[0023] Firstly, the hierarchical encoding that deeply integrates page layout and semantics avoids the problems of context fragmentation and loss of structural attributes caused by fixed window segmentation, significantly improving the accuracy of knowledge extraction.

[0024] Secondly, by using cross-document entity alignment and pseudo-co-occurrence verification, entity redundancy caused by synonymous writing is eliminated, ensuring the consistency and integrity of the knowledge graph.

[0025] Third, by using graph neural network association completion and rule-based soft priors, implicit associations between documents are automatically discovered, and interpretable association chains are output, making the completion results traceable and verifiable.

[0026] Fourth, by using a three-state governance state machine and active weight decay, the dynamic evolution and "soft deletion" of the knowledge graph are achieved, taking into account both the freshness of knowledge and the integrity of historical audit.

[0027] Fifth, by integrating graph and vector dual-channel retrieval with lineage permission management, interpretable intelligent retrieval, question answering, and impact analysis can be achieved, thereby improving the efficiency of knowledge reuse and the business's ability to withstand risks. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. 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. Wherein:

[0029] Figure 1 The method flowchart provided by the present invention. Detailed Implementation

[0030] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0031] This invention provides a knowledge asset management method based on artificial intelligence-based multi-source business document association analysis. Through hierarchical semantic parsing, cross-document entity alignment, relational reasoning, and dynamic evolution of knowledge graphs, it realizes the automatic transformation and continuous governance of business documents into structured knowledge assets, improves the accuracy of knowledge extraction, the completeness of association, and the level of intelligent management, thereby enhancing the efficiency of business decision-making and the ability to reuse knowledge.

[0032] For details, please refer to Figure 1 The present invention includes the following steps:

[0033] Step S1 – Acquisition and Unified Data Processing of Multi-Source Heterogeneous Documents: Original business documents from different sources and in varying formats are uniformly transformed into standardized text streams and document structure trees with layout information. This provides unbiased and complete data input for subsequent semantic analysis. Specifically, this includes the following input and processing flow:

[0034] Let the set of accessed data sources be D = {D1, D2, ..., D...} m}, where m is the number of data sources. Each data source D i Generate the original document set P i .

[0035] For each original document p, perform the following three steps:

[0036] (1) Format determination and unified conversion:

[0037] To determine the MIME type of p, use different channels as shown in Table 1:

[0038] Table 1: Mapping Table of Raw Format and Processing Channel

[0039]

[0040] After unified conversion, a standardized text line sequence L={l1, l2, ..., l...} is obtained. K}, where each text line l j Contains tuples:

[0041]

[0042] in:

[0043] text j The text content of the j-th line;

[0044] x j y j : The normalized horizontal and vertical coordinates (0 to 1) of the top left corner of this text line within the page;

[0045] w j h j: The normalized width and height of this text line;

[0046] page j Page number;

[0047] conf j : OCR confidence (constantly 1.0 for plain text extraction channels, used as a reliability marker for OCR channels).

[0048] (2) Page layout analysis and document structure tree modeling:

[0049] The text line clusters are structured and grouped using layout analysis models (such as CNN / Transformer-based face detection networks) to identify the types and geometric relationships of layout elements and construct a Document Structure Tree (DST).

[0050]

[0051] in:

[0052] N = {n1, n2, ..., n} nt}: A set of nodes on the page, each node being n u Represents a page element (title, subtitle, paragraph, table, block, header, footer, list item, etc.);

[0053] E⊆N×N: Set of parent-child hierarchical edges, (n u n v )∈E represents n u is n v The parent node (i.e., n) u Includes or governs n v ).

[0054] Each page node n u Having node attribute vectors:

[0055]

[0056] in:

[0057] type u Page element type tags (title / h2 / h3 / paragraph / table / figure / caption / list...);

[0058] level u : Hierarchical depth (0 at the root, +1 for each level below);

[0059] parent u Parent node index;

[0060] siblings u : The set of indices of sibling nodes at the same level;

[0061] sem textu : The text content contained in this node.

[0062] Through this step, the text line sequence L is organized into a tree structure T, with its root node representing the entire document and each level node accurately restoring the document's layout and logical hierarchy.

[0063] Step S2 – Document Semantic Parsing and Fragmentation Based on Page Layout Hierarchy Awareness: Based on the document structure tree, the document is segmented into semantically complete and hierarchical-aware "semantic chunks," and a hierarchical encoding vector that integrates structural context and semantic content is generated for each semantic chunk. This avoids context fragmentation and loss of hierarchical attributes caused by fixed-window segmentation. Specifically, it includes the following steps:

[0064] 2.1 Semantic block segmentation:

[0065] Starting from the leaf nodes of the document structure tree T, segmentation is performed according to semantic boundary determination rules. The segmentation decision is given by the semantic boundary classifier, which determines whether adjacent text units should be separated.

[0066] Let f be the feature vector of the adjacent block boundary. b The classifier outputs the disconnection probability P. split :

[0067]

[0068] in:

[0069] σ(): Sigmoid activation function, which maps the output to (0, 1);

[0070] W b : Segmentation classifier weight matrix (learnable);

[0071] b b : Segmentation classifier bias term (learnable);

[0072] f b ∈R df The boundary feature vector is composed of the following sub-features:

[0073] f b Sub-features include: (a) differences in layout type between two adjacent text units; (b) level difference |Δlevel|; (c) cosine similarity of topic vectors (topics are given by a lightweight topic model); and (d) punctuation / enumeration marker features.

[0074] When P split>τ b (τ) b The segmentation threshold is 0 < τ b When <1 (default 0.6), a split boundary is generated between the two elements.

[0075] After segmentation, we obtain a semantic block set C = {c1, c2, ..., c...} ∣C∣}, each semantic block c i At the same time, its corresponding page node chain is preserved:

[0076]

[0077] This indicates that the semantic block inherits from the hierarchical path from the root to the leaf in the node chain, thus preserving structured location information.

[0078] 2.2 Hierarchical-aware semantic encoding:

[0079] For each semantic block c i Perform three-stage hierarchical encoding to output a comprehensive semantic vector, including the following steps:

[0080] Step 1: Word-level encoding: Let semantic block c i The text was segmented into a token sequence w i，1 w i，2 , ..., w i，Li L i This represents the number of tokens in this block. Word-level encoding is performed using a pre-trained language model (such as BERT / RoBERTa) to obtain the hidden state of each word:

[0081]

[0082] in:

[0083] o i，t ∈R dplm : The hidden state vector of the t-th token after being processed by the pre-trained model;

[0084] d plm The hidden dimension of the pre-trained model (e.g., 768 for BERT-base).

[0085] Step 2: Sentence-level representation and block-level semantic vector: Perform weighted pooling on the hidden states of all tokens within the semantic block to obtain the plain text semantic vector:

[0086]

[0087] in:

[0088] t i ∈R dplm Semantic block c iPlain text semantic vectors;

[0089] α i，t ≥0, ∑ t α i，t =1: Attention weight, given by the attention scoring function.

[0090]

[0091] q: Learnable query vector (used to guide attention to important words);

[0092] W a Attention weight matrix (learnable).

[0093] Step 3: Structural Context Fusion (Hierarchical Awareness Fusion): Extracting the structural context vector s of semantic blocks. i The structural context consists of two parts:

[0094]

[0095] in:

[0096] v parent ∈R dplm Semantic block c i The semantic vector of the title of the parent node of the page node; if there is no title parent node, the average of all text vectors under that parent level is used.

[0097] N sib (i): Semantic block c i The set of sibling semantic block indexes (parallel blocks under the same parent node);

[0098] |N sib (i)∣: Number of sibling blocks, used to calculate the average;

[0099] j∈N sib (i) Traverse all sibling blocks, t j Its plain text semantic vector;

[0100] Concat[]: Vector concatenation operation, dimension stacking.

[0101] Finally, the text semantic vector and the structural context vector are mapped to a unified dimension through a gated fusion layer to obtain the final hierarchical semantic vector:

[0102]

[0103] in:

[0104] h i ∈R dh Semantic block c i The final hierarchical semantic vector, dh For the merged dimensions;

[0105] W cs ∈R dh×dcs : Structural context projection matrix (learnable), d cs For s-dimensional (d cs =2d plm );

[0106] ⊙: Element-wise vector multiplication (Hadamard product);

[0107] g i ∈(0,1) dh : Gated vector, d h As a dimension, it is obtained from the following formula:

[0108]

[0109] σ: Sigmoid activation, element-wise.

[0110] W g : Gating layer weight matrix (learnable), dimension R dh×(dplm+dcs) ;

[0111] b g Gating layer bias (learnable), dimension R dh .

[0112] When the structural context is highly relevant, g i →1, the model relies more on structural context; when the structure differs greatly from the text topic, g i →0, the model regresses to the semantics of the pure text. This achieves adaptive weight allocation between structure and semantics.

[0113] 2.3 Output: Each semantic block yields a final hierarchical semantic vector h. i Its structural path information is used as input features for subsequent extraction and alignment.

[0114] Step S3 – Fine-grained cross-document knowledge extraction and entity relationship annotation: For each semantic block of the generated hierarchical semantic vector, extract key business entities and semantic relationships between entities to provide basic triples for cross-document association and graph construction. This includes the following steps:

[0115] 3.1 Named Entity Recognition (NER): Employs a span-based nested entity recognition model to simultaneously identify ordinary entities and nested entities within a semantic block.

[0116] Let semantic block c i The content contains a token sequence. For each candidate span, the model outputs whether that span constitutes a certain type of entity.

[0117]

[0118] in:

[0119] span [p，q] The continuous text span from the p-th token to the q-th token;

[0120] o i，p o i，q The hidden state of the pre-trained model for the first and last tokens of this span;

[0121] h i Semantic block c i The hierarchical semantic vector (obtained from S2);

[0122] SpanFeat(): Span feature combination function;

[0123] W e Entity classification weight matrix (learnable);

[0124] softmax: Normalization, outputs the probability distribution of the span belonging to each entity type (such as project, amount, date, person in charge, terms, equipment code, location).

[0125] Expansion of span characteristic function:

[0126]

[0127] in:

[0128] (o i，q -o i，p ): Span: The difference vector between the first and last spans captures the overall semantic changes of the span;

[0129] P [p，q] : Span width position encoding (generated based on span length and starting offset);

[0130] After splicing, the entity type label is obtained by passing through a linear layer and softmax.

[0131] Incorporating domain priors: For entities such as amounts, dates, and codes, rule / dictionary post-processing is overlaid, and the rule extraction results are fused with the model results by confidence weighting to improve the accuracy of numerical entity extraction.

[0132] 3.2 Relation Extraction (Cross-Document Relation Recognition): Identify the semantic relationships between entity pairs using a prompt-based relation classification model.

[0133] Let there be a pair of entities (e) u e v), and construct a prompt template M(e) with the semantic block context. u e v c i ), input into the relation extraction model:

[0134]

[0135] in:

[0136] R: A predefined set of relationship types (such as "Contract-bound Project", "Solution Corresponding Project", "Meeting Decision-Making Person in Charge", "Mandatory Implementation of Clauses", etc.).

[0137] M(e u e v c i ): Place entity pairs into an input template composed of a semantic block context;

[0138] φ(): Template encoder (can be a PLM shared with S2 or a separate encoder), outputs a relational representation vector;

[0139] W r Relationship classification weight matrix (learnable);

[0140] : Predicted relation type label.

[0141] Extraction result triples: For each valid entity pair, generate a triple (head, rel, tail) and record its source document, semantic block and extraction confidence γ∈(0,1).

[0142] Step S4 – Cross-Document Entity Alignment and Global Knowledge Graph Construction: Entities describing the same business object in different documents are normalized into standardized nodes, eliminating synonyms and repetitions. Based on this, an enterprise-level global knowledge graph is constructed, specifically including the following steps:

[0143] 4.1 Multi-granularity entity alignment: For any two candidate entities e a (From document X) and e b (From document Y), calculate the similarity in three aspects:

[0144] (1) Attribute similarity:

[0145]

[0146] in:

[0147] K ab The set of attribute keys that appear in both entities (such as name, code, person in charge, amount, date, etc.);

[0148] |Kab |: Number of common attributes, used to calculate the average;

[0149] The value of the two entities on the k-th attribute;

[0150] 1{}: Indicator function, 1 when the two values ​​are approximately (exact match or fuzzy match), otherwise 0;

[0151] Edit distance normalized matching is used for named attributes, while relative tolerance matching is used for numerical attributes.

[0152] (2) Contextual semantic similarity: Using the semantic vectors obtained from S2, calculate the cosine similarity of the semantic block vectors of the two entities:

[0153]

[0154] in:

[0155] h c(a) Entity e a The semantic vector of the semantic block it belongs to;

[0156] h c(b) Entity e b The semantic vector of the semantic block it belongs to;

[0157] .: Vector dot product;

[0158] ∥∥: The L2 norm of a vector;

[0159] The result ranges from [-1, 1] to [-1, 1], where the closer to 1, the more similar the context.

[0160] (3) Structural similarity: Compare the hierarchical positions of the two entities in their respective document structure trees:

[0161]

[0162]

[0163] in:

[0164] level(a): Entity e a The level depth of the node in the DST;

[0165] type(a), type(b): The element types of the two entities on the page;

[0166] λ s >0: Structural difference attenuation coefficient (controlling the degree of influence of hierarchical differences);

[0167] β: Penalty coefficient for type inconsistency;

[0168] exp(): The exponential function that maps differences to similarities in the range (0, 1).

[0169] (4) Overall weighted similarity:

[0170]

[0171] in:

[0172] w1, w2, w3≥0 and w1+w2+w3=1: The weight coefficients (normalized) of the three types of similarity. They can be automatically learned by training logistic regression coefficients on the labeled development set, or they can be manually set (default 0.5 / 0.35 / 0.15, with the attribute weight being the largest).

[0173] When Sim(a, b) > τ merge When the merging threshold is set to 0.72, the two entities are determined to be the same object, and the merging process is executed.

[0174] (5) Pseudo-co-occurrence check (enhanced merge confidence): For confidence levels within the fuzzy interval (τ) low <Sim≤τ merge For candidate pairs of ), an additional pseudo-co-occurrence check is introduced:

[0175]

[0176] in:

[0177] Nucleus: The number of documents in which entity a and entity b appear simultaneously;

[0178] Denominator: The number of documents in which any of these entities appear;

[0179] Co(a, b) ∈ [0, 1]: co-occurrence rate, reflecting whether two entities co-occur repeatedly in multiple independent documents;

[0180] When Co(a, b) > τ co When the co-occurrence threshold is 0.6, the confidence level of merging is significantly improved, and the candidate can be included in the direct merging set or added to the high-confidence candidate pool.

[0181] After merging, a globally unique ID is assigned to each canonical entity, and an "alias table" is maintained (mapping the original names in each document to the canonical ID).

[0182] 4.3 Global Knowledge Graph Construction: All extracted and merged triples are written into the knowledge graph:

[0183]

[0184] in:

[0185] V: A set of canonical entity nodes (including alias mappings);

[0186] E: Set of relation edges, where each edge (u, v, r) represents an entity u being associated with entity v through relation type r;

[0187] R: Set of relation types;

[0188] Γ: The set of edge attributes, where each edge carries a triplet attribute. ,,in:

[0189] conf: Overall confidence score (composed of NER and RE confidence scores);

[0190] srcs: A collection of source documents and semantic block IDs (lineage);

[0191] t0: First write timestamp;

[0192] status: activity status (active / declining / ineffective), used for subsequent governance.

[0193] Step S5 – Graph Neural Network-Based Association Reasoning and Knowledge Completion: By learning entity embeddings through a graph neural network, combined with a translation distance model and rule-based soft priors, the system automatically completes the latent associations between documents that are not explicitly written, and outputs interpretable association paths to achieve "knowledge discovery." This includes the following steps:

[0194] 5.1 Graph Attention Network (GAT) Node Representation Learning: For a knowledge graph G, learn the structure-aware embedding of each entity:

[0195]

[0196] in:

[0197] : The embedding vector of entity u in the graph convolution at the k-th layer;

[0198] : The updated embedding vector;

[0199] N(u): The set of neighboring nodes of entity u;

[0200] W (k) : The learnable weight matrix of the k-th layer;

[0201] σ: Nonlinear activation (such as ReLU);

[0202] Attention coefficient, representing the importance weight of neighbor v to entity u, is calculated by the following formula:

[0203]

[0204] ∥∥: Vector concatenation;

[0205] a: Learnable attention vector;

[0206] LeakyReLU: ReLU activation with leak correction;

[0207] This formula performs softmax normalization on the neighbor weights.

[0208] The final entity embedding is the concatenation of the outputs from each layer:

[0209]

[0210] in:

[0211] K: Number of graph convolutional layers;

[0212] e u : The final embedding vector of entity u.

[0213] 5.2 Relationship Completion (Link Prediction): A translational distance model (TransE / RotatE) is used to model possible candidate triples. Taking RotatE as an example, entities and relations are modeled as rotations in complex space:

[0214] For a candidate triple(u, r, v), define the energy scoring function:

[0215]

[0216] in:

[0217] e u e v : Complex embedding vectors of head and tail entities (dimensions consistent with the graph).

[0218] r r : The complex plane rotation vector of relation r (with a magnitude of 1);

[0219] ∘: Element-wise product of complex vectors (represented as a rotation operation in RotatE, i.e., e^(-1 / 2)). u ∘r r (This is equivalent to embedding the head entity in a complex space and rotating it).

[0220] ∥∥2: L2 norm;

[0221] A higher score indicates a greater probability that the triple is true (a true triple should satisfy e). v ≈e u ∘r r ).

[0222] For any missing entity or relation, the candidate list is traversed and sorted by score to output the completion result:

[0223]

[0224] in:

[0225] The candidate with the highest score among the tail entities;

[0226] The set of all entities after deducting the head entity;

[0227] τ link Link completion score threshold (those below this threshold will not be accepted).

[0228] 5.3 Soft Prior Constraints: To ensure that the completion results conform to business logic, soft rule constraints are introduced to participate in the scoring. Let the rule set R be... rule Each rule R j Describe business constraints (such as "contracts under the same project must bind the project" or "the person in charge must belong to a certain organization").

[0229] The final completion was achieved through a joint scoring system:

[0230]

[0231] in:

[0232] μ∈(0,1): Prior weight coefficients of the rule;

[0233] 1{}: The newly added triple makes the knowledge graph satisfy rule R j When the condition is met, the reward is 1; when the condition is violated, the reward is 0... (measured by probability / consistency under soft constraints);

[0234] This mechanism ensures the generalization ability of graph neural networks while injecting business priors, thereby improving the adoption rate.

[0235] 5.4 Interpretable Association Path Generation: For the completed associations, a readable path explanation is generated. This employs path reasoning (weighted path enumeration):

[0236] Weights are transferred based on relation type, and a path connecting two entities is searched on the graph. Let the path P = (u→r1→e1→r2→⋯→e L-1 →r L →v), Path credibility:

[0237]

[0238] in:

[0239] L: Number of path steps (number of relation edges);

[0240] γ i : The overall confidence level of the relation edge at step i;

[0241] conf(e i ): Intermediate entity e i Extraction confidence level;

[0242] Return the highest-scoring paths with a limited number of steps as an interpretable chain of evidence for the user.

[0243] Step S6 – Dynamic Update of Knowledge Graph Based on Time-Sequence and Version Evolution: When business documents are added or changed, accurately identify the differences between the new information and the existing knowledge in the graph, and dynamically update the graph through a three-state governance state machine. This balances knowledge freshness and historical traceability, avoiding contamination by "dirty knowledge." Specifically, this includes:

[0244] 6.1 Document Version and Timeline Management: Maintain a version chain for each document. Each version p t Includes a timestamp t. New / modified documents first enter the incremental parsing channel to extract a new set of triples Δ.

[0245] 6.2 Incremental Difference Detection: For the extracted new triples (u, r, v) new Compare each item with the existing content of the map and calculate the degree of conflict:

[0246]

[0247] in:

[0248] Sim rel (r new r old ): Semantic similarity between old and new relation types (1 for the same relation, low value for different relations);

[0249] 1{}: Indication of whether an old triple exists with the same head and tail entities;

[0250] A high degree of conflict indicates that new information contradicts old knowledge.

[0251] 6.3 Three-state governance state machine: based on the new triplet confidence level γ new The governance action is determined by comparing the version's age with its relative newness. A comprehensive decision score is defined.

[0252]

[0253] in:

[0254] γ new New triplet extraction confidence level (NES / RE confidence level);

[0255] dist(t new , t old ): Time difference between old and new document versions;

[0256] λ d Time decay coefficient;

[0257] The decision segment Decide triggers different states in different intervals.

[0258] The state machine has three states, as shown in Table 2 below:

[0259] Table 2: State Machine State Statistics Table

[0260]

[0261] 6.4 Activity Weight Decay Mechanism: "Decaying" triples are assigned activity weights, which decay exponentially over time (soft deletion).

[0262]

[0263] in:

[0264] ω(t): The activity weight at the current time t;

[0265] ω0: Initial activity weight (normally 1);

[0266] t decline The moment when the state of decline begins;

[0267] η>0: Decay time constant (determines the decay rate);

[0268] When ω(t) decreases to the threshold τ ω (e.g., 0.1) or less, the triplet is removed from the "active search view" but is retained in the historical audit database and can be traced back.

[0269] 6.5 Conflict resolution and manual review: Candidates with high conflict levels and uncertain confidence levels are entered into the manual review queue, where domain experts confirm and then either write or reject them; manual operation records are entered into the audit log to ensure that the governance process is traceable.

[0270] Step S7 – Knowledge Assetization and Intelligent Application Services: Transforming the governed, consistent knowledge graph into searchable, question-answerable, reasoning-enabled, and auditable knowledge asset services to support business decisions. Specifically, this includes:

[0271] 7.1 Graph Vector Dual-Channel Fusion Retrieval: Construct two sets of parallel indexes, and merge and sort them during retrieval:

[0272] (1) Semantic vector channel

[0273] Map entities and semantic blocks to dense vectors eu (Reusing GAT embedding), the user query q is encoded as q v Calculate the cosine similarity:

[0274]

[0275] (2) Graph structure path channel: Perform structure retrieval on the knowledge graph and calculate the graph distance score between the entity and the query intent.

[0276] (3) Merging and sorting:

[0277]

[0278] in:

[0279] ρ∈[0,1]: The fusion weight of the vector channel and the graph channel (default 0.6, which can be tuned at the application layer);

[0280] Sim graph (q, e): Relevance score based on graph structure.

[0281] 7.2 Natural Language Knowledge Question Answering: User questions are parsed into graph queries by the intent recognition and entity linking module, solved on the knowledge graph subgraph, and an answer with a chain of evidence is generated.

[0282] Intent recognition: Determine whether it involves associative reasoning / constraint / impact analysis;

[0283] Entity linking: Aligns the references in the question to canonical entity nodes;

[0284] Subgraph solving: Perform constraint solving or path reasoning along the relational path to obtain candidate answers;

[0285] Evidence chain generation: Call the associated path generated in 5.4, attach the source document and page number to form verifiable evidence.

[0286] 7.3 Impact Analysis and Risk Warning: Given an entity (e.g., a project delay), deduce the affected downstream constraint nodes along the associated path, including contract terms, responsible persons, milestone dates, technical constraints, etc., and output an impact diffusion and transmission diagram. When a change in the state of a key triple (e.g., an important constraint entering the DECAY state) is detected, a risk assessment and warning are triggered.

[0287] 7.4 Access Control and Lineage Management: Each knowledge edge retains its srcs (source document + semantic block) lineage information, supporting one-click back to the original text; hierarchical sharing is implemented based on organizational roles and document security levels to prevent unauthorized access to sensitive knowledge; all extraction, update, attenuation, and manual review operations are written to the audit log to meet compliance requirements.

[0288] To better understand the technical solution of the present invention, the following specific embodiments are provided:

[0289] Example 1: Engineering Contract Management Scenario – Discovery of the Correlation between Contracts and Project Milestones:

[0290] Business Background: A construction company has recently received three new documents and needs to verify whether the system can automatically extract contractual constraints and complete implicit relationships through graph reasoning to output an interpretable chain of evidence.

[0291] Three input documents:

[0292] Document A - Contract: Construction Contract for "Huayu Road and Bridge Reconstruction Project", with a construction period constraint of "overall completion date of 2025-03-15".

[0293] Document B - Technical Solution: Technical solution document, the person in charge is "Li Jianhua", corresponding to "Huayu Road and Bridge Reconstruction Project".

[0294] Document C - Meeting Minutes: The meeting minutes state that "the milestone date for the Huayu Road and Bridge Project has been extended to 2025-05-20".

[0295]

Step S1

[0296] All three documents contain extractable text layers, and the text extraction channel is used with an OCR confidence level of conf=1.0. Based on layout analysis, a document structure tree is constructed. Taking document A as an example, the structure nodes are shown below:

[0297]

[0298] Calculation: The three documents generate a total of K=128 text lines, which, after slicing, result in a semantic block set |C|=36.

[0299]

Step S2

[0300] Take the semantic block c1 containing "Huayu Road and Bridge Reconstruction Project" as an example.

[0301] Word-level encoding (assuming d) plm =4 Simplified Demonstration): Assume the PLM outputs the hidden state of the 5 words in this block: o 1，1 =(0.2, 0.4, -0.1, 0.3), o 1，2 =(-0.1, 0.5, 0.2, 0.1),...,o 1，5 =(0.3, 0.2, 0.1, 0.4);

[0302] Weighted pooling: Assuming attention weights α = (0.1, 0.2, 0.15, 0.3, 0.25) (normalized to sum to 1), then the plain text semantic vector: t1 = 0.1o1，1 +0.2o 1，2 +0.15o 1，3 +0.3o 1，4 +0.25o 1，5 =(0.02, 0.04, -0.01, 0.03)+(-0.02, 0.10, 0.04, 0.02)+(0.045, 0.03, -0.015, 0.03)+(-0.015, 0.045, 0.045, 0.015)+(0.075, 0.05, 0.025, 0.10)=(0.105, 0.265, 0.085, 0.195);

[0303] Structural context: The parent node is the "Project Overview" title, and its vector v parent =(0.3, 0.1, 0.2, 0.5); the mean of the sibling block vectors in the same section is assumed to be (0.15, 0.2, 0.1, 0.1). Concatenating them, we get: s1=(0.3, 0.1, 0.2, 0.5, 0.15, 0.2, 0.1, 0.1);

[0304] Gating fusion: Let W g Simplified calculation yields g1 = (0.7, 0.6, 0.8, 0.5), W cs s1=(0.12, 0.08, 0.10, 0.14), then: h1=0.7(0.105, 0.265, 0.085, 0.195)+0.3(0.12, 0.08, 0.10, 0.14)=(0.0735, 0.1855, 0.0595, 0.1365)+(0.036, 0.024, 0.03, 0.042)=(0.1095, 0.2095, 0.0895, 0.1785);

[0305]

Step S3

[0306] Entity Recognition: After span NER and rule post-processing, the following entities were extracted:

[0307]

[0308] Relation extraction: Triples identified through prompt relation classification:

[0309]

[0310]

Step S4

[0311] We now need to determine whether e1 (Huayu Road and Bridge Renovation Project), e4 (Huayu Road and Bridge Renovation Project), and e5 (Huayu Road and Bridge Project) from different documents are the same entity.

[0312] Taking the alignment of e1 and e5 as an example:

[0313] Attribute similarity: Common attribute key K={name}, name "Huayu Road and Bridge Renovation Project" vs "Huayu Road and Bridge Project", edit distance normalized matching (assuming "renovation project / project" is identified as a synonymous suffix) yields a matching score of 0.85: Sim attr =0.85;

[0314] Contextual semantic similarity: the block containing e1 has a similarity h = (0.1095, 0.2095, 0.0895, 0.1785); the block containing e5 has a similarity vector of (0.12, 0.19, 0.10, 0.17). Therefore: Sim ctx =

[0315]

[0316]

[0317] (The cosine value is theoretically ≤1, rounded to 0.95).

[0318] Structural similarity: e1 and e5 both belong to the same depth of 2, both are of type "item description node", Δlevel=0, and their types are consistent, so we take λ. s =0.5, β=0.3: Sim struct =exp(-0.5×(0+0.3×0))=exp(0)=1.0;

[0319] Overall similarity (weight w=(0.5, 0.35, 0.15)): Sim(e1, e5)=0.5×0.85+0.35×0.95+0.15×1.0=0.425+0.3325+0.15=0.9075;

[0320] Because 0.9075 > τ merge =0.72, e1 and e5 are determined to be the same entity, and are merged.

[0321] Similarly, verify e1 and e4 (identical names, attribute similarity of 1.0, and highly similar context), Sim>0.72, and merge them.

[0322] Merge result: e1, e4, and e5 are merged into canonical entity V. project =Huayu Road and Bridge Reconstruction Project, with a confidence level of 0.95. Although e2 and e6 are both "dates", they have different semantic roles (one is the original completion date, and the other is the delayed date), so they are not merged.

[0323] Knowledge graph construction: Write the normalized triples to form the initial graph G0:

[0324]

Step S5

[0325] Graph attention networks learn the embeddings of each entity. Assume the embeddings are obtained through GAT propagation:

[0326] e(V project ) = (0.5, 0.3);

[0327] e(2025-03-15)=(0.4, 0.5);

[0328] e(Li Jianhua) = (0.6, 0.2);

[0329] e(2025-05-20)=(0.35, 0.45);

[0330] Missing Relationship Detection: The system found that there was a missing relationship edge between the node "Li Jianhua" and "2025-05-20". RotatE was used to complete the candidate relationship.

[0331] Let the rotation vector of the relation "responsible for acceptance" be r_acceptance = (e_t) iθ ), where θ is the rotation angle (take θ = 30°). Test candidate triplet "(Li Jianhua, responsible for acceptance, 2025-05-20)":

[0332] Let e 尾料 This indicates that "2025-05-20" is embedded, e 头 ∘r is approximately a 30° rotation of e(Li Jianhua) = (0.6, 0.2): (0.6, 0.2) → (0.6cos30° - 0.2sin30°, 0.6sin30° + 0.2cos30°) = (0.6 × 0.866 - 0.2 × 0.5, 0.6 × 0.5 + 0.2 × 0.866) = (0.4196, 0.4732);

[0333] Energy score: ;

[0334] The score is above the link threshold τ link =-0.5, and is better than other candidates, so complete the relationship "Li Jianhua → Responsible for acceptance → 2025-05-20".

[0335] Rule soft prior verification: Let rule R1 be: "After the project is delayed, the relevant person in charge should take charge of the acceptance of the new date node". The newly added triple satisfies R1. Let μ=0.1: FinalScore=-0.0733+0.1×1=0.0267>0, so the completion is adopted.

[0336] The system outputs the evidence chain for the associated path.

[0337]

Step S6

[0338] When document C (meeting minutes) is entered as a new version, a conflict is detected between "delayed to 2025-05-20" and the original "constrained completion date 2025-03-15" in the graph (two different completion dates for the same project).

[0339] Calculate the collision degree: Sim rel (Update constraint, constraint) = 0.4, the old triple exists: Conflict = (1 - 0.4) × 1 = 0.6;

[0340] Tristate determination: The confidence level of the new triplet is 0.89, and the time difference between the old and new triplets is t. new -t old (Assuming a difference of 90 days), take λ. d =0.002:Decide=0.89-0.002×90=0.89-0.18=0.71;

[0341] 0.71>τ upd =0.5, Execute update state: The newly added "Constraint Completion 2025-05-20" is in an active state, and the original "Constraint Completion 2025-03-15" is in a decaying state. Active weight: When η=60, after 30 days of decline: ω=exp(-30 / 60)=exp(-0.5)≈0.607, which then decreases to τ. ω When the value is below 0.1, the original date is removed from the active view but the history is retained.

[0342]

Step S7

[0343] Impact Analysis: A user asked, "If the project is delayed, which constraints will be affected?" The system deduced the impact along the path in the graph.

[0344] Output the impact transmission diagram and issue a risk warning for "Li Jianhua" regarding "change in acceptance deadline".

[0345] Dual-channel retrieval: For the query "Who is responsible for the Huayu Road and Bridge Project", the semantic vector channel score is 0.92 and the graph structure channel score is 0.88. Taking ρ=0.6: Rank=0.6×0.92+0.4×0.88=0.552+0.352=0.904.

[0346] Example 2: Financial Credit Review Scenario – Cross-validation and Correlation Risk Identification:

[0347] Business Background: Bank customer service teams need to review loan applications, and the system needs to automatically link the borrower's loan contract, financial report, and credit record documents to identify hidden credit risks.

[0348] Three input documents:

[0349] Document D - Loan Contract: Lender: "Zhang Wei", loan amount: "5 million", term: "24 months".

[0350] Document E - Financial Report: Annual report of the borrowing enterprise "Zhang Wei Technology Co., Ltd.", which states "current assets 20 million, current liabilities 18 million".

[0351] Document F - Credit Report: It shows that "Zhang Mou" applied for three loans, two were approved, and there is an overdue record.

[0352] [Step S1] Unified digitization:

[0353] Documents D and E are extractable texts, and Document F is a scanned document processed through the OCR channel. Taking Document F as an example, the OCR output of a certain page includes confidence: l1=(...if the confidence of recognizing the character "Zhang Wei" is 0.85...) Suppose a total of K=210 text lines are generated, and the number of semantic blocks |C|=58.

[0354] [Step S2] Hierarchical encoding:

[0355] Taking the financial data block c3 of Document E as an example, a vector (4-dimensional for demonstration) is obtained through hierarchical gating encoding: h3=(0.31, 0.42, 0.18, 0.55);

[0356] The features are embodied in the structural context of "financial terms + report chapters".

[0357] [Step S3] Entity and relation extraction:

[0358]

[0359] Relation extraction: identify triples:

[0360] (E1, obtain loan, E2): confidence 0.88;

[0361] (E1, loan term, E3): confidence 0.91;

[0362] (E4, as borrower, ——): alignment confirmation required;

[0363] (E6, has overdue, ——): alignment required;

[0364] [Step S4] Entity alignment:

[0365] The system needs to judge: Are E1 (Zhang Wei) and E5 (Zhang Mou), E1 and E4 (Zhang Wei Technology Co., Ltd.) the same entity?

[0366] Align E1 (Zhang Wei) and E5 (Zhang Mou):

[0367] Attribute similarity: Name "Zhang Wei" vs "Zhang Mou", edit distance normalized = 0.8 (matching the same part of the surname).

[0368] Context similarity: The contract block vector of E1 is h=(0.4, 0.2, 0.1, 0.3), the credit block vector of E5 is h=(0.41, 0.19, 0.12, 0.29), Sim ctx =0.97.

[0369] Structural similarity: Both appear at the "borrower" level, Sim struct =0.9.

[0370] Overall similarity (w=(0.5, 0.35, 0.15): Sim=0.5×0.8+0.35×0.97+0.15×0.9=0.4+0.3395+0.135=0.8745.

[0371] Since 0.8745 > 0.72, E1 and E5 are determined to be the same person and are merged into the canonical entity V. 张伟 .

[0372] Align E1 (Zhang Wei) and E4 (Zhang Wei Technology Company):

[0373] Attribute similarity: Name "Zhang Wei" vs "Zhang Wei Technology Company", low match score of 0.45 (the company entity and the individual are different).

[0374] Context similarity: Personal loan context vs. corporate financial context, Sim ctx =0.30.

[0375] Structural similarity: Different hierarchical types (individual vs. enterprise), Sim struct =0.4.

[0376] Sim=0.5×0.45+0.35×0.30+0.15×0.4=0.225+0.105+0.06=0.39.

[0377] 0.39 < 0.72, not merged — correctly determines that the individual borrower and the enterprise are different entities, but an "affiliation" (such as a holding / employment relationship) needs to be established between the two.

[0378] Based on this connection, the system further identifies a "legal representative" relationship between E1 and E4 and writes it into the graph.

[0379]

Step S5

[0380] Key point to note: Financial institutions often pay attention to whether "the same controlling shareholder has risks in multiple places".

[0381] Calculate corporate credit association:

[0382] The GAT calculation yields an embedding, and the candidate relationship "has an overdue association".

[0383] Weighted path: .

[0384] Assumption completion (V) 伟科技公司 (There are hidden risks, overdue records) : Score=-0.28>τ link =-0.5 is adopted, forming a hidden risk association.

[0385] Soft prior rule: Suppose the rule "The credit records of enterprises and individuals with the same legal representative should be checked together", μ=0.15: FinalScore=-0.28+0.15=-0.13<0?

[0386] This is a negative value. The system resets the "high weight, high risk" rule to μ=0.5: FinalScore=-0.28+0.5×1=0.22>0, which is adopted, but marked as requiring manual review (entering the manual review queue in S6).

[0387]

Step S6

[0388] The key point here is that the "overdue" information comes from the newly inserted credit report document F. The "new triplet" (V) 张伟 There is an overdue period (E6) and there is no direct conflict with the old graph. Conflict=0, Decide=0.90 (high confidence new information). Execute the new state, write to the legal node and make it active.

[0389] Meanwhile, new edges were added to the graph for subsequent impact analysis.

[0390]

Step S7

[0391] The question, "Does Zhang Wei have any historical risks?", is answered "yes" by the system integrating semantics and graph paths. The evidence chain shows: Contract D, a loan of 5 million → Corporate financial leverage (current ratio = current assets / current liabilities = 2000 / 1800 ≈ 1.11, which is low) → Credit delinquency. The risk level is output as "moderate to high," and manual review is recommended.

[0392] Dual-channel fusion:

[0393] The semantic channel score is 0.85 ("Zhang Wei risk" semantic match);

[0394] Image channel 0.90 (path integrity);

[0395] Rank=0.6×0.85+0.4×0.9=0.51+0.36=0.87;

[0396] Meanwhile, financial ratio calculations serve as an additional business logic to support decision-making.

[0397] Example 3: Policy Interpretation Scenario – Policy Consistency Analysis Across Documents:

[0398] Business Background: Government agencies or large organizations need to compare policy documents issued by multiple departments, identify inherent conflicts or synergies between policy requirements, and provide compliance impact analysis.

[0399] Three input documents:

[0400] Document G - Environmental Policy: "Regulations on the Prevention and Control of Air Pollution in a Certain City" (Article 5: New industrial projects must obtain environmental impact assessment permits before construction).

[0401] Document H - Investment Promotion Policy: "Preferential Measures for Investment Promotion in a Certain City" (Article 3: Major projects can be "built first and approved later" for rapid implementation).

[0402] Document I - Project Implementation Report: Implementation report of "a certain industrial project", which records that the project has started construction, but the environmental impact assessment has not been completed and approved.

[0403]

Step S1

[0404] All three texts can be directly extracted to generate conf=1.0. A total of K=98 text lines and |C|=27 semantic blocks are generated.

[0405]

Step S2

[0406] The paragraph block c5 containing Article 5 of the Environmental Protection Regulations is encoded as: h5=(0.45, 0.30, 0.20, 0.55), (the structural context includes "Article 5 / New Industrial Projects");

[0407] Article 3 of the investment promotion policy is located in block c3: h3=(0.2, 0.1, 0.5, 0.3), (the structural context includes "Article 3 / Build first, then approve");

[0408]

Step S3

[0409]

[0410] Relational triples:

[0411] (P2, Requirements, P1): 0.90 (Environmental impact assessment is required for new industrial projects);

[0412] (P3, Permitted, P4): 0.87 (Major projects can be built first and approved later);

[0413] (P5, belongs to, P2): 0.93 (a certain industrial project belongs to a newly built industrial project);

[0414] (P5, Status, P6): 0.92 (Environmental impact assessment not yet completed during project implementation);

[0415]

Step S4

[0416] Align P2 (new industrial projects) and P3 (major investment projects): It is necessary to determine whether they are of the same category and need to be associated.

[0417] Attribute similarity: Project category "New Industrial Project" vs "Major Project", match 0.55.

[0418] Context similarity: Environmental protection regulations context vs. investment promotion policy context, Simctx=0.4 Simctx=0.4.

[0419] Structural similarity: Significant hierarchical differences (environmental enforcement vs. investment promotion), Sim struct =0.5.

[0420] Sim=0.5×0.55+0.35×0.4+0.15×0.5=0.275+0.14+0.075=0.49<0.72.

[0421] Instead of merging, establish a "subclass-parent class" association:

[0422] Major investment projects ⊂ Project category, which may overlap with "new industrial projects";

[0423] Align P5 (an industrial project) with P2 (a newly established industrial project): Link specific projects of entity P5 to category P2 to form a hierarchy:

[0424] P5 (a certain industrial project) -- belongs to --> P2 (new industrial project) -- requires --> P1 (environmental impact assessment).

[0425]

Step S5

[0426] Core reasoning: To examine whether there is a contradiction between "building first and then approving" and "pre-environmental impact assessment".

[0427] In the diagram:

[0428] P5 (a certain industrial project) --belongs to --> P2 (a newly established industrial project);

[0429] P2 --Requirement--> P1 (Pre-conference Environmental Impact Assessment);

[0430] P5 -- Status --> P6 (Environmental impact assessment not completed);

[0431] Meanwhile, the investment promotion policy P3 (major projects) allows for "construction first, approval later." The project report shows that "a certain industrial project is already under construction."

[0432] Soft prior rule: Rule R2 states: "When a pre-approval requirement conflicts with a post-approval operation, it should be upgraded to a violation risk."

[0433] Graph Reasoning Scoring: A new triplet has been added: "(An industrial project has a risk of non-compliance, and the environmental impact assessment has not been completed)". Score = 0.75 > τ link Add a rule R2 reward, μ=0.2: FinalScore=0.75+0.2×1=0.95. High confidence adoption is marked as high risk.

[0434]

Step S6

[0435] Project Report I is a new document. Its "Environmental Impact Assessment Incomplete" is newly added information, with Conflict=0, and is written in a new state. The system also marks the entity to which it belongs in the "Candidate Set of Violations", pending policy review or manual confirmation.

[0436]

Step S7

[0437] Impact Analysis: When a user asks, "Are there any policy-related approval obstacles to a certain industrial project?", the system outputs the result along the graph.

[0438] Dual-channel search:

[0439] Semantic channel: The query "obstacles to approval of a certain industrial project" yields 0.80;

[0440] Image channel: Path integrity 0.88;

[0441] Rank=0.6×0.80+0.4×0.88=0.48+0.352=0.832;

[0442] The system outputs a compliance analysis summary and evidence chain, which can be used by policy implementers to assess whether the environmental protection supplementary approval process needs to be initiated or whether the "build first, approve later" policy exemption should be applied, thus realizing intelligent policy compliance management based on "data-driven + rule-assisted".

[0443] 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 asset management method based on artificial intelligence-driven multi-source business document correlation analysis, characterized in that: Includes the following steps: Step 1: Perform unified data processing on multi-source heterogeneous business documents. Extractable text, scanned image, table-intensive, and rich media hybrid documents are processed through text extraction, OCR recognition, table parsing, layout analysis, and OCR combined channels to construct a document structure tree. Step 2: Based on the document structure tree, the document is segmented into semantic blocks carrying structural context using a semantic boundary classifier, and then fused with word-level encoding, sentence-level weighted pooling and structural context gating to generate hierarchical-aware semantic vectors; Step 3: Based on hierarchical semantic vectors, extract key business entities through a span-level named entity recognition model, identify semantic relationships between entities through a prompt relationship extraction model, and generate knowledge triples with confidence and source lineage. Step 4: Combine attribute similarity, contextual semantic similarity, and structural similarity, and combine pseudo-co-occurrence verification to normalize and align cross-document entities, and construct a global knowledge graph; Step 5: Based on the entity structure embedding learned by the graph attention network, the knowledge graph is completed by combining the translation distance model and the rule-based soft prior, generating interpretable association paths; Step 6: Detect conflicts between new and old knowledge based on document version and time sequence, execute knowledge addition, update or decay operations through polymorphic governance state machine, and achieve soft deletion by decaying the activity weight, so as to dynamically maintain the knowledge graph; Step 7: Based on the governed knowledge graph, semantic association retrieval and knowledge question answering are provided through a dual-channel fusion index of graph vectors, and hierarchical sharing is implemented according to lineage and permissions.

2. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 1, characterized in that, Step two generates hierarchical semantic vectors, including: constructing a document structure tree T=(N, E), and extracting structural context vectors s for each semantic block. i Transform the plain text semantic vector t i With s i Hierarchical semantic vectors are generated through the gated fusion layer. , where the gate vector W cs W g b g is a learnable parameter, and ⊙ represents element-wise multiplication.

3. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 2, characterized in that, The pseudo co-occurrence verification in step four includes: for candidate entity pairs whose overall similarity is within the fuzzy range, calculating their co-occurrence rate Co(a, b) in multiple independent documents; when the co-occurrence rate is higher than the preset co-occurrence threshold, increasing the merging confidence of the candidate entity pair and performing merging.

4. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 3, characterized in that, Step 5, relation completion, includes: modeling triples in the complex embedding space using rotations, traversing candidate entities according to their energy scoring function to complete the completion; and using business rules as soft priors, giving a reward score to correct the completion result when a new triple makes the graph satisfy the rules.

5. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 4, characterized in that, Step five generates interpretable association paths, including: searching for paths connecting two entities on the knowledge graph through weighted path reasoning; the confidence of a path is represented by the product of the confidence of each relation edge and the confidence of the intermediate entity; and returning the path with the highest score and a limited number of steps as the evidence chain.

6. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 5, characterized in that, Step six, the polymorphic governance state machine, includes: when there is no corresponding old triplet in the graph and the comprehensive decision score is higher than the addition threshold, perform addition; when there is a conflicting old triplet and the decision score is higher than the update threshold, perform update and retain the historical version chain; when old information is replaced or the confidence decreases, it is set to a decay state, and the activity weight is given to decay exponentially over time to achieve soft deletion.

7. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 6, characterized in that, Comprehensive decision score Calculate, where γ new For the new triple confidence level, λ d Here, dist(t) is the time decay coefficient. new , t old () represents the time difference between the old and new versions.

8. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 7, characterized in that, Step seven involves dual-channel fusion of graph vectors, including: merging the cosine correlation of the semantic vector channel with the graph distance correlation of the graph structure channel. The fusion sort is used, where ρ∈[0,1] is the fusion weight.

9. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 8, characterized in that, Step seven also includes: generating an evidence chain containing source documents and page numbers from the knowledge question-and-answer results; deducing the affected downstream constraint nodes along the associated path and outputting an impact diffusion transmission diagram, triggering a risk warning when the state of key knowledge changes.

10. The knowledge asset management method based on artificial intelligence multi-source business document association analysis according to claim 9, characterized in that, All extraction, alignment, completion, update, decay, and manual review operations generated in steps one through seven are recorded as audit logs, and each knowledge edge retains the lineage information of the source document and semantic block, supporting the tracing back to the original text and hierarchical sharing.

Images