A feedback-enhanced knowledge base and knowledge graph bidirectional coupling method

Abstract

The application relates to a feedback enhancement-based knowledge base and knowledge graph bidirectional coupling method, which comprises the following steps: forward structured modeling; graph-driven reverse generation; cross-layer confidence coupling propagation; time sequence sensing-based relationship weight self-adaptive updating; change detection-based self-evolution triggering; consistency constraint and checking. The application has the beneficial effect that the application converts knowledge graph reasoning results into natural language knowledge items to fill the knowledge base through a reverse completion channel, and realizes the real bidirectional circulation between the two types of knowledge carriers.

Description

Technical Field

[0001] This invention relates to the field of knowledge engineering technology, and more specifically, to a bidirectional coupling method for knowledge bases and knowledge graphs based on feedback enhancement. Background Technology

[0002] Knowledge bases and knowledge graphs are two core forms of knowledge management systems. Knowledge bases store knowledge in the form of documents, question-answer pairs, or structured records, suitable for carrying complete semantic context and narrative logic, facilitating human understanding and direct retrieval. Knowledge graphs represent knowledge using entity-relation-entity triples (h, r, t), supporting multi-hop reasoning, relation chain analysis, and graph structure mining. In practical knowledge-intensive applications (such as intelligent question answering, decision support, and scientific literature management), these two are often deployed separately, unaware of each other, leading to the following series of technical problems:

[0003] Problem 1: The single-path forward flow leads to low knowledge utilization. Existing methods typically treat the knowledge base only as the raw data source for the knowledge graph. Triples are extracted from text and written into the graph through an information extraction pipeline, after which the knowledge base and the graph remain independent and unrelated. A large number of implicit relationships discovered by the knowledge graph through path reasoning and structure mining (such as "A is a superordinate concept of B and instances of B have attribute P, therefore instances of A may also have attribute P") cannot be fed back into the knowledge base, resulting in a serious waste of implicit knowledge. Existing knowledge graph enhancement systems generally lack a reverse path from the graph to the knowledge source, which is one of the main bottlenecks restricting the overall effectiveness of the knowledge system.

[0004] Problem 2: Knowledge consistency cannot be guaranteed. Original documents in knowledge bases often suffer from issues such as conflicting statements, outdated information, and heterogeneous sources. Existing methods lack cross-document confidence aggregation mechanisms when constructing knowledge graphs, making it impossible to assess the overall credibility of the same entity relationship across different sources. Furthermore, the structural information of the knowledge graph (such as node centrality and subgraph density) is not used to back-evaluate the quality of knowledge base documents, resulting in the graph being filled with noisy triples with low confidence, thus polluting the reasoning process of downstream applications.

[0005] Question 3: Lack of a collaborative mechanism for dynamic updates. In real-world scenarios, the document content of the knowledge base continuously grows and is revised, requiring the knowledge graph to be dynamically updated accordingly. Existing incremental update methods (such as FastKGE and AIR methods) focus on efficiently updating the graph embedding when new triples arrive, but they do not consider how changes in the graph structure (such as a significant increase in the weight of a relation) should reversely adjust the knowledge base's retrieval index and document ranking. Furthermore, they have not established a collaborative evolution framework that connects new documents → graph updates → knowledge base reordering into a closed loop, and have not addressed the architectural design of bidirectional collaborative evolution.

[0006] Question 4: The temporality of relationships is overlooked. The timeliness of knowledge is one of the core attributes of a knowledge system. Most of the relationship weights in existing knowledge graphs are static values, which cannot reflect the intuition that "relationships that have been repeatedly confirmed recently should be more reliable than those that were rarely mentioned in the past".

[0007] In summary, existing technologies lack a unified architecture that can tightly couple knowledge bases and knowledge graphs, enable bidirectional flow, cross-layer confidence propagation, relation weight time-series awareness, and self-evolution triggering. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bidirectional coupling method for knowledge bases and knowledge graphs based on feedback enhancement.

[0009] Firstly, a feedback-enhanced bidirectional coupling method for knowledge bases and knowledge graphs is provided, including:

[0010] S1. Forward structured modeling steps: Read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple.

[0011] S2, Graph-driven reverse generation steps: Perform structural reasoning on the knowledge graph to deduce implicit relation triples, and use the graph-aware text generation model to reversely transform the derived new triples into natural language description paragraphs, and write the paragraphs back to the knowledge base in the form of knowledge entries.

[0012] S3, Cross-layer confidence coupling propagation stage: Propagate confidence signals between the document layer, triplet layer and graph structure layer, and synchronize the confidence of the three layers through a mapping function;

[0013] S4, Time-aware Relationship Weight Adaptive Update Stage: Time-aware weights are assigned to relation edges in the knowledge graph. A composite time-series model including time decay and event incentives is used to dynamically update the weights. When the weight change exceeds a preset threshold, the knowledge base index is adjusted.

[0014] S5. Self-evolution triggering stage based on change detection: Monitor the new document flow of the knowledge base, and when a new document containing new knowledge is detected, automatically trigger the local update process of steps S1 to S4.

[0015] S6. Consistency Constraint and Verification Phase: Perform consistency verification on the knowledge entries that are reverse-filled in step S2. Entries that fail verification are marked as pending review and isolated.

[0016] Preferably, in S1, the initial confidence score is obtained by weighted summation of the posterior probability output by the extraction model, the semantic similarity between the head and tail entities, and the normalized frequency value of the relation in the document.

[0017] Preferably, the graph-aware text generation model in step S2 includes:

[0018] The structure encoding module is used to encode the neighborhood subgraph centered on the target triple using a graph neural network, and output a context-aware entity representation;

[0019] The semantic template selection module is used to select a narrative template from the candidate template library based on the relation type and entity category of the target triple;

[0020] The text generation module is used to generate a natural language description from a language model based on the graph encoding output and the narrative template.

[0021] The discriminative filter is used to calculate the semantic similarity between the generated text and existing similar entries in the knowledge base. If the similarity exceeds a threshold, regeneration is triggered.

[0022] Preferably, the mapping function in S3 includes:

[0023] Mapping from document layer to triplet layer: After linearly transforming the document confidence and extracted confidence using a learnable weight matrix, the triplet confidence is output through an activation function;

[0024] Mapping from triplet layer to graph structure layer: Node confidence is updated through iterative propagation. In each iteration, the confidence of a node is determined by the weighted sum of its own confidence and the confidence of its neighboring nodes.

[0025] Feedback mapping from graph structure layer to document layer: After the confidence score of the graph structure layer converges, the corresponding document is multiplied by the corresponding gain coefficient according to the different intervals in which the graph confidence score falls, and the document confidence score is updated.

[0026] Preferably, in S4, the composite time series model includes an exponential decay term and an accumulation term based on historical event stimuli; the exponential decay term is determined by the base weight, decay rate, and time interval since the last update, and the event stimuli term is obtained by accumulating the confidence changes at each historical moment after decay through a kernel function; the index adjustment is triggered when the comprehensive weight is lower than the lower threshold or higher than the upper threshold.

[0027] Preferably, the knowledge change detector described in S5 employs a difference graph mechanism, specifically including:

[0028] Construct a local temporary graph for the newly added documents;

[0029] Calculate the difference graph between the local temporary graph and the existing knowledge graph, and extract the set of newly added triples and the set of conflicting triples;

[0030] Perform local embedding incremental update on the triples in the newly added triple set, updating only the entity embeddings of the affected subgraphs;

[0031] A conflict resolution strategy is applied to the triples in the set of conflicting triples.

[0032] Preferably, the local embedding incremental update employs an incremental adapter mechanism, including:

[0033] Dynamic rank allocation: Based on the structural characteristics and confidence changes of each entity in the new triplet in the difference graph, the adapter rank of each entity is dynamically allocated. The greater the confidence change, the higher the rank is allocated.

[0034] Cross-layer parameter isolation: The affected subgraph is divided into multiple layers according to its topological distance from the existing graph. Adapter updates are performed only on entity embeddings in some layers, while parameters in other layers are frozen.

[0035] Coupled-aware fine-tuning: A knowledge base feedback regularization term is introduced into the loss function of the adapter. The feedback regularization term is based on the pass rate of the backfilled text after consistency verification, so that the graph embedding update and knowledge base quality feedback form a closed loop.

[0036] Preferably, the method also includes an online learning optimization step: continuously collecting user feedback signals on the knowledge base retrieval results, converting the feedback signals into training samples, and using them to fine-tune the information extraction model in step S1 and the text generation model in step S2.

[0037] Secondly, a feedback-enhanced bidirectional coupling system of knowledge base and knowledge graph is provided for performing any of the methods described in the first aspect, including:

[0038] The forward extraction module is used to read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple.

[0039] The reverse generation module is used to perform structural reasoning on the knowledge graph to deduce implicit relation triples. The derived new triples are then converted into natural language description paragraphs through the graph-aware text generation model, and the paragraphs are written back to the knowledge base as knowledge entries.

[0040] The confidence propagation module is used to propagate confidence signals between the document layer, triplet layer, and graph structure layer, and synchronizes the confidence levels of the three layers through a mapping function.

[0041] The weight update module is used to assign time-aware weights to the relation edges in the knowledge graph. It uses a composite time-series model that includes time decay and event incentives to dynamically update the weights. When the weight change exceeds a preset threshold, it triggers an adjustment of the knowledge base index.

[0042] The self-evolution module is used to monitor the flow of new documents in the knowledge base. When a new document containing new knowledge is detected, a partial update process is automatically triggered.

[0043] The verification module is used to perform consistency verification on the knowledge entries that are filled in backwards. Entries that fail verification are marked as pending review and isolated.

[0044] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0045] The beneficial effects of this invention are:

[0046] 1. Breaking through the bottleneck of one-way flow at the data utilization level of the knowledge retrieval system: This invention transforms the results of knowledge graph reasoning into natural language knowledge entries to fill the knowledge base through a reverse completion channel, realizing true two-way flow between the two types of knowledge carriers, increasing the amount of retrievalable implicit knowledge in the knowledge base by 30% to 50% (estimated based on simulation experiments); at the same time, since implicit knowledge is directly entered into the database in the form of natural language, the frequency of downstream applications calling the graph reasoning service during retrieval is reduced, thereby reducing the real-time computing load and network transmission overhead of the system.

[0047] 2. At the level of data storage and quality management, the consistency of global knowledge is improved: The cross-layer confidence propagation mechanism of this invention correlates the quality signals of the document layer, triple layer and graph structure layer, so that the triple noise rate of the knowledge graph and the proportion of low-quality documents in the knowledge base can be reduced at the same time; accordingly, the amount of conflicting data that the system needs to process during the data cleaning stage is reduced, the redundancy of the knowledge base index is reduced, and the effective utilization of storage resources is improved.

[0048] 3. At the system operation and maintenance and update scheduling level, automated collaborative evolution is achieved: The self-evolution controller of this invention connects the update processes of the two types of knowledge carriers into an automatically triggered closed loop, which can maintain the timeliness of the knowledge system without manual intervention; through the differential graph mechanism and local embedding incremental update (IncLoRA), the waste of CPU and GPU computing resources and system service interruption caused by traditional full retraining are avoided, and the update cycle of large-scale knowledge graphs is significantly shortened.

[0049] 4. Enhance knowledge time-series awareness in time-sensitive business systems: The composite time-series model of this invention enables relation weights to sensitively reflect changes in the timeliness of knowledge. When relation weights change significantly, it automatically triggers a local reordering of the knowledge base index, allowing the search engine to prioritize returning documents with higher timeliness and reducing the number of invalid queries caused by users obtaining outdated information.

[0050] 5. In terms of data security and system reliability, the reverse knowledge security is guaranteed: The consistency verifier of this invention performs entity pointer verification, logical consistency check and temporal legality verification on the reverse-filled content, keeping the risk of illusion pollution within a manageable range; the pre-filtering mechanism reduces the probability of unqualified data entering the knowledge base, reduces the backlog length of the subsequent manual review queue and system I / O overhead, and meets the security requirements of high-reliability knowledge service scenarios.

[0051] 6. At the distributed system architecture level, it has engineering scalability: The event-driven message bus architecture of this invention enables each functional module to be independently and horizontally scaled, supporting the online service deployment of large-scale knowledge systems; each module is decoupled through asynchronous message queues, allowing the knowledge base retrieval engine and the knowledge graph storage engine to dynamically adjust computing nodes and memory resources according to their respective loads, thereby improving cluster resource utilization and concurrent processing capabilities. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the overall architecture of the feedback-enhanced bidirectional coupling system of knowledge base and knowledge graph of the present invention;

[0053] Figure 2 This is a schematic diagram of the internal process of the forward extraction channel;

[0054] Figure 3 This is a schematic diagram of the internal process of the reverse completion channel;

[0055] Figure 4 A schematic diagram of the cross-layer confidence propagation mechanism;

[0056] Figure 5 This is a schematic diagram of the time series curve for the dynamic adjustment module of relation weights.

[0057] Figure 6 A schematic diagram of the differential graph mechanism of the self-evolving controller;

[0058] Figure 7 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0059] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0060] Example 1:

[0061] The following problems exist in the existing technology:

[0062] (1) There is only a one-way forward flow channel between the knowledge base and the knowledge graph. The reasoning results of the graph cannot be used to fill the knowledge base in reverse, resulting in low utilization of implicit knowledge.

[0063] (2) There is a lack of a unified confidence propagation mechanism across the three knowledge levels of document layer, triplet layer and graph structure layer, which makes it impossible to guarantee knowledge consistency;

[0064] (3) Knowledge base updates and knowledge graph updates are independent of each other, lacking a collaborative triggering and closed-loop evolution mechanism;

[0065] (4) The relation weights are static values ​​and cannot reflect the temporal changes of knowledge;

[0066] (5) There is a lack of consistency verification mechanism for the reverse-filled content, which poses a risk of illusory knowledge polluting the knowledge base.

[0067] To address the problems of existing technologies, Embodiment 1 of this application provides a bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement, such as... Figure 7 As shown, Figure 7 The text uses rectangles to represent each processing step, diamonds to represent decision nodes (such as "whether a new entity / relationship was detected" and "whether the consistency verification passed"), and arrows to indicate the direction of control flow, fully presenting the execution logic and loop of the six main steps.

[0068] Specifically, the method includes:

[0069] S1. Forward structured modeling steps: Read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple.

[0070] In S1, the initial confidence score is obtained by weighted summation of the posterior probability output by the extraction model, the semantic similarity between the head and tail entities, and the normalized frequency value of the relation in the document. Each weight coefficient is obtained through training or empirical setting.

[0071] Specifically, the initial confidence score is calculated as follows:

[0072]

[0073] These are the weighting coefficients, and . For the head entity and tail entity, the embedding vectors are... For cosine similarity, This is the normalized frequency value of relation r in the document.

[0074] S2, Graph-driven reverse generation steps: Perform structural reasoning on the knowledge graph to deduce implicit relation triples, use a graph-aware text generation model to reverse-engineer the derived new triples into natural language description paragraphs, and write the paragraphs back to the knowledge base in the form of knowledge entries.

[0075] The graph-aware text generation model described in S2 adopts a joint architecture based on graph attention networks and pre-trained language models, including:

[0076] (a) Graph Neural Network-based Encoding Module: Relational Graph Convolutional Network (R-GCN) is used to encode the K-hop subgraph centered on the target triple, and output a context-aware entity representation;

[0077] (b) Semantic template selection module based on relation type: Selects the optimal narrative template from the candidate template library according to the relation type and entity category of the target triple, wherein the template library is pre-classified and stored according to relation ontology;

[0078] (c) Text generation module based on pre-trained language model: Natural language description is generated by autoregression of the decoder of the pre-trained language model, using the graph encoding output of the structure encoding module and the narrative template of the semantic template selection module as conditional inputs.

[0079] (d) Discrimination filter: Calculate the semantic similarity between the generated text and existing similar entries in the knowledge base. If the similarity exceeds the preset threshold, regeneration is triggered to ensure the diversity of the reverse-filled content.

[0080] S3, Cross-layer confidence coupling propagation stage: Propagate confidence signals between the document layer, triplet layer and graph structure layer, and synchronize the confidence of the three layers through a mapping function.

[0081] In S3, confidence synchronization between the three levels is achieved through the following mapping function:

[0082] (a) Mapping from document layer to triplet layer:

[0083]

[0084] Where σ is the Sigmoid activation function, and W1 and W2 are learnable weight matrices;

[0085] (b) Mapping from triplet layer to graph structure layer: In the belief propagation iteration, the confidence of node v in the k-th iteration is:

[0086]

[0087] For nodes The layer confidence score in the k-th iteration. The self-confidence retention coefficient is used. For nodes The set of neighboring nodes, This represents the attention weights of neighboring nodes.

[0088] (c) Feedback mapping from the graph structure layer to the document layer: After the confidence score of the graph structure layer converges, the confidence score of the corresponding document is updated through the inverse mapping function, triggering the review and annotation of low-confidence documents in the knowledge base. The formula for updating the confidence score is as follows:

[0089]

[0090] For the updated document confidence level, The mapping gain function is as follows:

[0091]

[0092] S4, Time-aware Relationship Weight Adaptive Update Stage: Time-aware weights are assigned to relation edges in the knowledge graph. A composite time-series model including time decay and event incentives is used to dynamically update the weights. When the weight change exceeds a preset threshold, the knowledge base index is adjusted.

[0093] In S4, the composite time series model includes an exponential decay term and an accumulation term based on historical event incentives; the exponential decay term is determined by the base weight, decay rate, and time interval since the last update, and the event incentive term is obtained by accumulating the confidence changes at each historical moment after decay through a kernel function; the index adjustment is triggered when the comprehensive weight is lower than the lower threshold or higher than the upper threshold.

[0094] Specifically, the relation weight update considers both time decay and the reinforcing effect of historical events, causing the relation weights to change dynamically over time and respond to new knowledge input. The weight update formula for the composite time series model described in step S4 is as follows:

[0095]

[0096]

[0097]

[0098] in, The basic weights of relation edge e, The decay rate over time. This is the time of the last overall update. For Hawkes incentive terms, Update time for historical events For incentive intensity, It is an exponential kernel function. for Change in confidence level over time.

[0099] When W(e,t) < Or W(e,t)> When the index rearrangement signal is triggered, wherein The lower threshold is... This is the upper limit threshold.

[0100] S5. Self-evolution triggering stage based on change detection: Monitor the new document flow in the knowledge base. When a new document containing new knowledge is detected, the local update process of steps S1 to S4 is automatically triggered.

[0101] The knowledge change detector described in S5 employs a difference graph mechanism, specifically including:

[0102] (a) Construct a local temporary graph G_new for the newly added document;

[0103] (b) Calculate the difference graph ΔG = G_new ⊕ G_existing between G_new and the existing knowledge graph G_existing, and extract the set of newly added triples T_add and the set of conflicting triples T_conflict;

[0104] (c) Perform local embedding incremental updates on the triples in T_add, updating only the entity embeddings of the affected subgraph. These local embedding incremental updates employ an improved incremental low-rank adapter mechanism, the improvements including:

[0105] (c1) Dynamic rank allocation based on confidence gradient: The adapter rank r_i of each entity is dynamically calculated based on the degree of its neighboring nodes and the change in confidence in the difference graph of each entity in the newly added triplet, where r_i = f(|Δconf|), and f is a monotonically increasing function.

[0106]

[0107] High-confidence changes are assigned to nodes with higher degree counts and larger ranks, while low-confidence changes are assigned to nodes with lower degree counts and smaller ranks, so that the rank assignment matches the intensity of knowledge changes.

[0108] (c2) Cross-layer parameter isolation strategy: The affected subgraph is divided into near-end layer, middle layer and far-end layer according to the topological distance with the existing graph. Low-rank adaptation update is performed only on the entity embedding of the near-end layer and the middle layer, and the parameters of the far-end layer are completely frozen to avoid catastrophic forgetting.

[0109] (c3) Two-way coupling perception fine-tuning: Introduce the knowledge base feedback regularization term L_feedback= λ × (1 - pass_ratio) into the adapter loss function, where pass_ratio is the pass rate of the most recent backfilled text after S6 consistency verification. The feedback signal is the pass ratio of the backfilled text after consistency verification. Use this ratio as the regularization coefficient to adjust the learning rate of the adapter, so that the graph embedding update and knowledge base quality feedback form a closed loop.

[0110] (d) Implement conflict resolution strategies for the triples in T_conflict, and determine the retained version according to the three-level arbitration rule of "timeliness first, confidence second, source authority verification".

[0111] S6. Consistency Constraint and Verification Phase: Perform consistency verification on the knowledge entries that are reverse-filled in step S2. Entries that fail verification are marked as pending review and isolated.

[0112] Specifically, it checks whether all entities mentioned in the reverse-filled text have corresponding nodes in the knowledge graph entity set; for entities that do not exist, it triggers the entity linking module to perform candidate matching. If the matching confidence is lower than the threshold θ, the entire entry is marked as pending review to avoid phantom entities polluting the knowledge base.

[0113] Example 2:

[0114] Based on Example 1, Example 2 of this application provides a more specific method for bidirectional coupling of knowledge base and knowledge graph based on feedback enhancement.

[0115] This embodiment is deployed on the internal knowledge management platform of a research institution. The knowledge base stores the full text and abstracts of approximately 150,000 scientific and technological documents (knowledge base size: approximately 200GB of raw text), and the knowledge graph contains approximately 850,000 entity nodes and approximately 6.2 million relation triples (covering three disciplines: materials science, biomedicine, and computer science). The system uses Elasticsearch as the knowledge base retrieval engine, Neo4j as the knowledge graph storage engine, and Apache Kafka as the event-driven message bus.

[0116] Specifically, the method provided in this application embodiment constructs a cross-layer coupling closed-loop control mechanism between the document layer, the triplet layer, and the graph structure layer, including:

[0117] S1. Forward structured modeling steps: Read the original document from the knowledge base and perform named entity recognition, relation extraction, and event detection on the original document; during the extraction process, identify and skip paragraphs carrying knowledge graph reasoning source tags, and only perform triple extraction on the original source document; write the extracted entity set, relation triple set, and event set into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple.

[0118] For example, for newly added scientific and technological documents, the system first performs document segmentation (by paragraph or fixed token window), and then calls a multi-task joint extraction model based on RoBERTa-large (NER annotation head + relation classification head + event detection head sharing a 12-layer Transformer encoder). In the field of materials science, the NER tag system includes four types of entities: MATERIAL, PROPERTY, APPLICATION, and PROCESS; relation types include 16 categories such as has_property, used_in, synthesized_by, and derived_from.

[0119] An example of initial confidence calculation is as follows: For the triple (zinc oxide nanorod, has_property, wide bandgap semiconductor), let P_extract = 0.82 (model posterior probability), cos_sim(e_h, e_t) = 0.61 (entity embedding cosine similarity), IDF_score(has_property) = 0.45 (has_property relations are relatively frequent in the corpus, and a lower value is taken after normalization. Here, IDF_score represents the scarcity of the relation, and the frequency is negatively correlated with it. The lower the value, the higher the confidence). Let α=0.6, β=0.25, γ=0.15, then conf_extract = 0.6×0.82 + 0.25×0.61 + 0.15×0.45 = 0.492 + 0.153 + 0.068 = 0.713.

[0120] The confidence score is written to Neo4j along with the triple and stored in the confidence attribute field of the relation.

[0121] S2. Graph-driven reverse generation steps: Perform structural reasoning on the knowledge graph, derive implicit relation triples using a path ranking algorithm and a rule-based reasoning engine, and then reverse-convert the derived new triples into natural language description paragraphs using a graph-aware text generation model; embed machine-recognizable source tags in the natural language description paragraphs to identify that the paragraphs are knowledge graph reasoning generated content rather than the original document; write the natural language description paragraphs carrying source tags back to the knowledge base in the form of knowledge entries to complete the reverse filling of the knowledge base.

[0122] For example, the system triggers a batch reverse completion task once a day at a set time. The reverse completion process is as follows:

[0123] (a) Path Ranking Inference: The system executes a Path Ranking Algorithm (PRA) based on random walks in Neo4j, inferring paths with an enumeration length ≤ 3 for entity pairs with a confidence score higher than 0.75. For example, if there are (Material A, has_property, Wide Bandgap) and (Wide Bandgap, enables, UV photocatalysis) in the graph, PRA derives the implicit relation (Material A, potentially_enables, UV photocatalysis), and the inference confidence score is the square root (i.e., the geometric mean) of the product of the confidence scores of the two edges.

[0124] (b) R-GCN subgraph encoding: Centered on the target triple (A material, potentially_enables, UV photocatalysis), extract a 2-hop neighborhood subgraph (about 30-50 nodes), pass messages through 2 layers of R-GCN, and output the context-aware embedding vector of the entity (dimension 256).

[0125] (c) Template Planning and Language Model Generation: The template set corresponding to the `potentially_enables` relation in the template library is queried, and the template with the highest confidence is selected. "Research shows that {head entities} have the potential to {relational descriptions} and {evidence path descriptions}, indicating that they may have significant application value in the {tail entities} domain." Subsequently, the domain-fine-tuned Qwen3.5-4B language model (fine-tuned using domain literature instructions) generates the following knowledge entry text using subgraph encoding and the template as conditions:

[0126] Studies have shown that zinc oxide nanorods (ZnO nanorods) possess potential for ultraviolet photocatalysis applications. Experimental results demonstrate that zinc oxide nanorods exhibit wide bandgap semiconductor properties, and wide bandgap materials have been proven to possess excellent photocatalytic activity under ultraviolet light activation, indicating their potential significant application value in the field of ultraviolet photocatalysis. [Source: KG Inference, Confidence Level: 0.71, Generation Time: 2026-04-07]

[0127] (d) Differentiated filtering: Calculate the BM25+ semantic dual-path recall score between the generated text and existing similar documents in Elasticsearch. If the semantic similarity with any existing document exceeds 0.88, regeneration is triggered (up to 3 times). Otherwise, the text is written to the knowledge base in a pending review state.

[0128] S3, Cross-layer Confidence Coupling Propagation Stage: Confidence signals are propagated between the document layer, triplet layer, and graph structure layer. The document layer confidence is determined by the information density score and the time decay coefficient; the triplet layer confidence is obtained by weighted summation of the posterior probability of the extraction model and the entity neighborhood aggregation value; the graph structure layer confidence is iteratively updated on the graph by the belief propagation algorithm until convergence; the three layers are synchronized with each other through a mapping function.

[0129] Specifically, the system performs confidence level synchronization on a daily batch processing cycle. Taking a scientific document d1 (original confidence level conf_doc(d1) = 0.85) as an example:

[0130] Document layer → triplet layer mapping: For the triplet extracted from d1 (silicate glass, has_property, high transmittance), conf_triple = σ(W1×0.85 + W2×0.71 + b1) = σ(0.83) ≈ 0.697 (W1, W2, b1 are parameters obtained from training).

[0131] Triple layer → Graph structure layer mapping: In the k-th iteration of belief propagation, the layer confidence of node "Silicate Glass" is updated as follows: conf_graph^(k)("Silicate Glass") = 0.4×0.697 + 0.6×[0.3×conf_graph^(k-1)("Quartz Glass") + 0.3×conf_graph^(k-1)("Optical Glass") + 0.4×conf_graph^(k-1)("Solar Cell")]. After 5 iterations, the confidence converges to 0.712.

[0132] Graph structure layer → document layer feedback: If the final layer confidence of the "Silicate Glass" node is higher than 0.70, then the retrieval weight of all documents from this field in the knowledge base will be increased, and the index weight of document d1 will be updated from 0.85 to 0.85 × 1.05 = 0.893 (the gain coefficient of 1.05 is determined by the mapping function).

[0133] The calculation formula is as follows:

[0134] For the updated document confidence level, The mapping gain function is as follows:

[0135]

[0136] S4. Time-aware relation weight adaptive update stage: Each relation edge in the knowledge graph is assigned a time-aware weight, and a composite time-series model including time decay and burst reinforcement terms is used to update the weight. When a new document arrives in the knowledge base, a local weight recalculation is triggered. When the change in relation weight is detected to exceed a preset bidirectional threshold, an index reordering signal is sent to the knowledge base. The bidirectional threshold includes two cases: the weight falls below the lower limit threshold and the weight exceeds the upper limit threshold.

[0137] For example, for the relation edge e1 = (ZnO, has_property, wide bandgap), let its base weight w_base = 0.73, the last update time be Δt = 30 days ago, λ = 0.005 (material domain time decay rate), the two most recent update times be t1 = 20 days ago (Δconf = +0.08) and t2 = 5 days ago (Δconf = +0.12), μ = 0.2, κ(Δt) = exp(-0.1×Δt):

[0138] w(e1, t) = 0.73 × exp(-0.005×30) + 0.2×0.08×exp(-0.1×20) + 0.2×0.12×exp(-0.1×5) = 0.73 × 0.861 + 0.016 × 0.135 + 0.024 × 0.607 = 0.629+ 0.002 + 0.015 = 0.646

[0139] Since θ_max > 0.646 > θ_min, where θ_min = 0.5 and θ_max = 0.9, the edge weight remains valid. If the weight of a relation falls below θ_min, a KG_WEIGHT_DROP event is sent to the Kafka message queue. After the knowledge base index service subscribes to this event, it reduces the retrieval priority of related documents.

[0140] S5. Self-evolution triggering stage based on change detection: Continuously monitor the new document flow of the knowledge base. When an incremental document is determined by the knowledge change detector to contain new entities or new relationships, the local update process of steps S1 to S4 is automatically triggered to complete the co-evolution of the knowledge graph and the knowledge base.

[0141] For example, when a batch of new documents (such as 200 new papers added daily) enter the knowledge base, the self-evolutionary controller executes the following process:

[0142] (a) Perform incremental information extraction on 200 new documents and construct a temporary graph G_new (containing the set of newly extracted triples).

[0143] (b) Calculate the difference graph ΔG by comparing hash fingerprints: T_add (new triples) about 2800, T_conflict (triples that conflict with the existing graph) about 140.

[0144] (c) Perform local embedding incremental updates on the triples in T_add, identify the affected subgraph (a 2-hop neighborhood centered on the newly added triples, involving approximately 120,000 nodes), and perform improved incremental low-rank adapter (IncLoRA) fine-tuning only on the embedding vectors of these 120,000 nodes:

[0145] ① Dynamic Rank Allocation Based on Confidence Gradient: Introducing the confidence change Δconf as a dynamic adjustment factor, and setting the rank r_i = 32 sigmoid(|Δconf|), r_i needs to be rounded up, where sigmoid(x) is a monotonically increasing function:

[0146]

[0147] When an entity’s confidence level changes significantly due to the injection of new documents, it automatically acquires a larger adapter rank to fully learn new knowledge; otherwise, it compresses the rank size to reduce parameter overhead.

[0148] ② Cross-layer parameter isolation strategy: The adapter is enabled only for the near layer within 1 hop and the mid layer within 2 hops of the newly added triplet, while the parameters of the far layer beyond 2 hops are completely frozen.

[0149] ③ Two-way Coupling Awareness Fine-tuning: A knowledge base feedback regularization term L_feedback = λ × (1 - pass_ratio) is introduced into the adapter loss function, where pass_ratio is the pass rate of the most recent backfilled text after S6 consistency verification. When the pass rate decreases, the weight of the regularization term is increased to suppress aggressive updates of the adapter and prevent noisy knowledge in new documents from polluting the graph through embedding updates; when the pass rate increases, the weight of the regularization term is decreased to accelerate the absorption of new knowledge. This mechanism creates a closed-loop coupling between graph embedding updates and knowledge base quality.

[0150] (d) For the 140 conflicting triples in T_conflict, the following three-level arbitration rules are applied: ① Timeliness arbitration: Compare the publication dates of the original documents corresponding to the two triples. If the time difference is >2 years, the newer one is retained; ② Confidence arbitration: If the time difference is ≤2 years, compare the confidence levels of the two triples and retain the one with higher confidence; ③ Source authority: If the difference in confidence levels is <0.05, the version to be retained is determined based on the impact factor (IF) of the journal from which the document is sourced.

[0151] S6. Consistency Constraint and Verification Phase: Entity pointer verification, logical consistency check, temporal legality check, and source cycle detection are performed on the knowledge entries that are reverse-filled in step S2. The source cycle detection is used to identify whether the entry has been reverse-generated by the knowledge graph. If a cycle generation trace is detected, the entry is rejected. Entries that fail verification are marked as pending manual review and isolated from the formal storage of the knowledge base.

[0152] For example, perform triple validation on each batch of inverted text (approximately 350 lines):

[0153] ① Entity reference verification: Use the entity linking service based on the BLINK model to check the entity references in each text. If the entity link confidence is <0.60, the entire record is marked as PENDING_REVIEW;

[0154] ② Logical consistency check: Compare statements involving numerical ranges (such as "the melting point of material X is Y℃") with the attribute values ​​already stored in the knowledge graph. If the difference exceeds 20%, a conflict alarm is triggered.

[0155] ③ Timing validity check: For statements containing temporal relationships (such as "A occurs before B"), check the logical consistency by comparing them with the existing temporal triples in the graph to prevent temporal reversal due to reasoning errors.

[0156] The verification pass rate was approximately 91.3% (based on internal test set statistics), and 8.7% of the items awaiting review were pushed to the manual review queue. Only after manual confirmation could they be officially added to the database.

[0157] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0158] Example 3:

[0159] Based on Embodiment 2, Embodiment 3 of this application provides a bidirectional coupling system of knowledge base and knowledge graph based on feedback enhancement, such as... Figure 1 As shown, Figure 1The diagram shows that the knowledge base (left) and knowledge graph (right) are interconnected through a forward extraction channel (solid arrow, from left to right) and a backward completion channel (dashed arrow, from right to left). An event-driven message bus (horizontal bidirectional arrow) runs through the middle of the two. The confidence propagation layer and the relation weight dynamic adjustment layer are located above the bus, while the consistency verifier and the self-evolution controller are located below the bus.

[0160] Specifically, the system includes:

[0161] The forward extraction module is used to read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple.

[0162] The forward extraction module employs a multi-task joint model, sharing underlying encoding among Named Entity Recognition (NER), Relation Extraction (RE), and Event Detection (ED) to reduce computational overhead. Each extracted triple is accompanied by an initial confidence score calculated based on a weighted average of model posterior probability, entity semantic similarity, and relation type rarity, providing a starting point for subsequent confidence propagation. Figure 2 As shown, Figure 2 It demonstrates the complete processing flow from the original document input, through document preprocessing, multi-task joint model (NER / RE / ED shared encoder), triple candidate generation, initial confidence calculation, graph pattern matching, and finally writing into the knowledge graph.

[0163] The reverse generation module is used to perform structural reasoning on the knowledge graph to deduce implicit relation triples. The derived new triples are then converted into natural language description paragraphs through a graph-aware text generation model, and the paragraphs are written back to the knowledge base as knowledge entries.

[0164] The reverse generation module derives implicit relationships from the knowledge graph using the Path Ranking Algorithm (PRA) and a rule mining engine. Then, a graph-aware text generation model converts the derived results into natural language text and writes it back to the knowledge base. The graph-aware text generation model employs a three-stage architecture: R-GCN encoding of subgraph structure information, a template planner selecting narrative strategies, and a pre-trained language model executing conditional generation. It is also equipped with a discriminative filter to prevent generated content from duplicating existing entries.

[0165] like Figure 3 As shown, Figure 3 The document demonstrates the processing flow from the knowledge graph reasoning module (path ranking algorithm + rule engine), through implicit triple generation, subgraph encoding (R-GCN), template planning and selection, language model condition generation, discriminative filtering, consistency verification, and finally writing back to the knowledge base.

[0166] The confidence propagation module is used to propagate confidence signals between the document layer, triplet layer, and graph structure layer, and synchronizes the confidence levels of the three layers through a mapping function.

[0167] The confidence propagation module establishes a top-down propagation path from the document layer to the triple layer to the graph structure layer, and a bottom-up feedback path from the graph structure layer to the document layer. Through a bidirectional mapping function, confidence signals are synchronized between the three layers to achieve a bidirectional consistency constraint that "document quality affects graph confidence, and graph confidence feeds back into document scoring".

[0168] like Figure 4 As shown, Figure 4 The diagram uses a three-layer structure to represent the document layer, triple layer, and graph layer. Solid arrows from top to bottom represent confidence downlink mapping, while dashed arrows from bottom to top represent confidence feedback mapping. Belief propagation (BP) iteration is represented by bidirectional arrow loops within the graph layer.

[0169] The weight update module is used to assign time-aware weights to the relation edges in the knowledge graph. It uses a composite time-series model that includes time decay and event incentives to dynamically update the weights. When the weight change exceeds a preset threshold, it triggers an adjustment of the knowledge base index.

[0170] The weight update module introduces a composite time-series model that includes a time decay term and a Hawkes activation term. It models two types of time-series effects at the knowledge graph relation weight level: "recently occurring dense relations are more reliable" and "confidence mutations should propagate rapidly." It uses significant changes in relation weights as trigger signals to drive local reordering of the knowledge base index.

[0171] like Figure 5 As shown, Figure 5 The horizontal axis represents time t, and the vertical axis represents the relation weight w(e,t). It shows the combined weight curve after the time decay curve (exponential downward trend) and the Hawkes incentive term are superimposed, as well as the annotation time when the weight falls below the threshold θ and triggers the knowledge base index reordering.

[0172] The self-evolution module is used to monitor the flow of new documents in the knowledge base. When a new document containing new knowledge is detected, a partial update process is automatically triggered.

[0173] The self-evolution module uses a differential graph mechanism to monitor the incremental document flow. By extracting the set of newly added triples and the set of conflicting triples, it triggers local embedding incremental updates (using IncLoRA technology to avoid catastrophic forgetting) and three-level arbitration conflict resolution, respectively, connecting "new document → graph update → knowledge base reordering" into a complete self-evolution closed loop.

[0174] like Figure 6 As shown, Figure 6 The document demonstrates the process of incremental document stream input, temporary graph construction, difference graph computation (Δ G = G_new ⊕ G_existing), branching between the newly added triple set T_add and the conflicting triple set T_conflict, and finally triggering local embedding updates and conflict resolution.

[0175] The verification module is used to perform consistency verification on the knowledge entries that are filled in backwards. Entries that fail verification are marked as pending review and isolated.

[0176] The verification module performs entity pointer verification, logical consistency check, and temporal legality check on the knowledge entries that are filled in backwards, to prevent illusory entities and logical contradictions from polluting the knowledge base and to ensure the overall credibility of the symbiotic system.

[0177] The above six modules work together through an event-driven message bus to form a closed "symbiotic system".

[0178] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

Claims

1. A bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement, characterized in that, include: S1. Forward structured modeling steps: Read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple. S2, Graph-driven reverse generation steps: Perform structural reasoning on the knowledge graph to deduce implicit relation triples, and use the graph-aware text generation model to reversely transform the derived new triples into natural language description paragraphs, and write the paragraphs back to the knowledge base in the form of knowledge entries. S3, Cross-layer confidence coupling propagation stage: Propagate confidence signals between the document layer, triplet layer and graph structure layer, and synchronize the confidence of the three layers through a mapping function; S4, Time-aware Relationship Weight Adaptive Update Stage: Time-aware weights are assigned to relation edges in the knowledge graph. A composite time-series model including time decay and event incentives is used to dynamically update the weights. When the weight change exceeds a preset threshold, the knowledge base index is adjusted. S5. Self-evolution triggering stage based on change detection: Monitor the new document flow of the knowledge base, and when a new document containing new knowledge is detected, automatically trigger the local update process of steps S1 to S4. S6. Consistency Constraint and Verification Phase: Perform consistency verification on the knowledge entries that are reverse-filled in step S2. Entries that fail verification are marked as pending review and isolated.

2. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 1, characterized in that, In S1, the initial confidence score is obtained by weighted summation of the posterior probability output by the extraction model, the semantic similarity between the head and tail entities, and the normalized frequency value of the relation in the document.

3. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 2, characterized in that, The graph-aware text generation model mentioned in step S2 includes: The structure encoding module is used to encode the neighborhood subgraph centered on the target triple using a graph neural network, and output a context-aware entity representation; The semantic template selection module is used to select a narrative template from the candidate template library based on the relation type and entity category of the target triple; The text generation module is used to generate a natural language description from a language model based on the graph encoding output and the narrative template. The discriminative filter is used to calculate the semantic similarity between the generated text and existing similar entries in the knowledge base. If the similarity exceeds a threshold, regeneration is triggered.

4. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 3, characterized in that, The mapping functions described in S3 include: Mapping from document layer to triplet layer: After linearly transforming the document confidence and extracted confidence using a learnable weight matrix, the triplet confidence is output through an activation function; Mapping from triplet layer to graph structure layer: Node confidence is updated through iterative propagation. In each iteration, the confidence of a node is determined by the weighted sum of its own confidence and the confidence of its neighboring nodes. Feedback mapping from graph structure layer to document layer: After the confidence score of the graph structure layer converges, the corresponding document is multiplied by the corresponding gain coefficient according to the different intervals in which the graph confidence score falls, and the document confidence score is updated.

5. The bidirectional coupling method for knowledge base and knowledge graph based on feedback enhancement according to claim 4, characterized in that, In S4, the composite time series model includes an exponential decay term and an accumulation term based on historical event stimuli. The exponential decay term is determined by the basic weights, decay rate, and time interval since the last update. The event stimuli term is obtained by accumulating the confidence changes at each historical moment after decay through a kernel function. The index adjustment is triggered when the overall weight is lower than the lower threshold or higher than the upper threshold.

6. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 5, characterized in that, The knowledge change detector described in S5 employs a difference graph mechanism, specifically including: Construct a local temporary graph for the newly added documents; Calculate the difference graph between the local temporary graph and the existing knowledge graph, and extract the set of newly added triples and the set of conflicting triples; Perform local embedding incremental update on the triples in the newly added triple set, updating only the entity embeddings of the affected subgraphs; A conflict resolution strategy is applied to the triples in the set of conflicting triples.

7. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 6, characterized in that, The local embedded incremental update adopts an incremental adapter mechanism, including: Dynamic rank allocation: Based on the structural characteristics and confidence changes of each entity in the new triplet in the difference graph, the adapter rank of each entity is dynamically allocated. The greater the confidence change, the higher the rank is allocated. Cross-layer parameter isolation: The affected subgraph is divided into multiple layers according to its topological distance from the existing graph. Adapter updates are performed only on entity embeddings in some layers, while parameters in other layers are frozen. Coupled-aware fine-tuning: A knowledge base feedback regularization term is introduced into the loss function of the adapter. The feedback regularization term is based on the pass rate of the backfilled text after consistency verification, so that the graph embedding update and knowledge base quality feedback form a closed loop.

8. The bidirectional coupling method between a knowledge base and a knowledge graph based on feedback enhancement as described in claim 7, characterized in that, It also includes an online learning optimization step: continuously collecting user feedback signals on knowledge base retrieval results, converting the feedback signals into training samples, and using them to fine-tune the information extraction model in step S1 and the text generation model in step S2.

9. A bidirectional coupling system of knowledge base and knowledge graph based on feedback enhancement, characterized in that, For performing the method as described in any one of claims 1 to 8, comprising: The forward extraction module is used to read the original document from the knowledge base, extract information from the original document, write the extracted entities, relations and events into the knowledge graph according to the predetermined graph pattern, and attach an initial confidence score to each triple. The reverse generation module is used to perform structural reasoning on the knowledge graph to deduce implicit relation triples. The derived new triples are then converted into natural language description paragraphs through the graph-aware text generation model, and the paragraphs are written back to the knowledge base as knowledge entries. The confidence propagation module is used to propagate confidence signals between the document layer, triplet layer, and graph structure layer, and synchronizes the confidence levels of the three layers through a mapping function. The weight update module is used to assign time-aware weights to the relation edges in the knowledge graph. It uses a composite time-series model that includes time decay and event incentives to dynamically update the weights. When the weight change exceeds a preset threshold, it triggers an adjustment of the knowledge base index. The self-evolution module is used to monitor the flow of new documents in the knowledge base. When a new document containing new knowledge is detected, a partial update process is automatically triggered. The verification module is used to perform consistency verification on the knowledge entries that are filled in backwards. Entries that fail verification are marked as pending review and isolated.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 8.

Images