Multi-source heterogeneous knowledge fusion question and answer solving system

Abstract

The invention discloses a multi-source heterogeneous knowledge fusion question and answer solution system, which relates to the technical field of computers and comprises a knowledge base layer, a recall layer, an analysis and pruning layer and an answer generation layer. The knowledge base layer is used for constructing a dynamic heterogeneous entity fusion engine and a self-adaptive knowledge slice storage mechanism, and the dynamic heterogeneous entity fusion engine comprises a cross-modal entity alignment algorithm, relation topology completion and an incremental entity evolution model; the self-adaptive knowledge slice storage mechanism is used for performing semantic density perception slicing on the RAG document and constructing a three-level index tree; according to the method, through the technologies of cross-modal entity alignment, dynamic recall weight adjustment, context sensing entity priority correction, RAG cross-block reasoning enhancement, double-encoder conflict detection, knowledge graph guide generation and the like, knowledge fusion reasoning is achieved, the complex question answering accuracy is effectively improved, the conflict recognition rate is increased, and the manual maintenance cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a multi-source heterogeneous knowledge fusion question-answering system. Background Technology

[0002] In the field of domain knowledge question answering technology, the current mainstream approach is to leverage large models to implement question answering functionality. However, the effective application of large models relies heavily on the infusion of domain knowledge. Existing knowledge bases can be mainly divided into three categories: FAQ question-answer pair knowledge bases, KG knowledge graphs, and RAG document knowledge bases. These three types of knowledge bases exhibit significant heterogeneity: FAQs are semi-structured knowledge bases, existing in the form of question-answer pairs; KGs are knowledge graphs constructed from structured knowledge; and RAGs target unstructured document knowledge, with their knowledge base being an unstructured vector knowledge base.

[0003] These different types of knowledge bases are suitable for different scenarios and each has its own advantages and disadvantages. However, a single technical approach based on a single knowledge base can no longer meet actual needs and is gradually becoming outdated. Currently, most question-answering systems adopt an independent recall method based on multi-source heterogeneous knowledge bases, which places extremely high demands on question distribution. But in reality, the boundaries of many questions are quite blurred, which makes question distribution extremely difficult.

[0004] Furthermore, single knowledge bases rely on either entity recall or semantic vector recall, both of which have significant drawbacks, such as inaccurate recall results and the potential for omissions. Moreover, complex problems often require comprehensive reasoning based on knowledge from multiple sources, a requirement that independent recall technologies simply cannot meet. It is against this technological backdrop that "a multi-source heterogeneous knowledge fusion question-answering system" has emerged, aiming to address the numerous problems existing in the aforementioned technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-source heterogeneous knowledge fusion question-answering system to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides a multi-source heterogeneous knowledge fusion question-answering system, comprising: Knowledge base layer: used to build a dynamic heterogeneous entity fusion engine and an adaptive knowledge slicing storage mechanism. The dynamic heterogeneous entity fusion engine includes cross-modal entity alignment algorithms, relation topology completion and incremental entity evolution models. The adaptive knowledge slicing storage mechanism is used to perform semantic density-aware slicing of RAG documents and build a three-level index tree. Recall layer: used to perform dynamic adjustment of recall weights in mixed modalities, dynamic correction of entity priority hierarchy, and cross-block inference enhancement for RAG fine ranking; Parsing and pruning layer: used to implement dynamic semantic threshold adjustment and knowledge conflict detection and resolution. Knowledge conflict detection and resolution includes a semantic-level conflict recognition engine, dynamic evaluation of knowledge source credibility, and a hierarchical execution strategy for conflict resolution. Answer generation layer: Used to guide the generation of hierarchical answers based on knowledge graphs.

[0007] Furthermore, the cross-modal entity alignment algorithm adopts a three-level alignment mechanism, including semantic anchor mapping, relation topology completion, and incremental entity evolution model. Semantic anchor mapping uses a large model to generate entity semantic fingerprints and calculates cross-database entity similarity through an attention mechanism.

[0008] Furthermore, the dynamic adjustment of mixed-modal recall weights is introduced into a reinforcement learning framework, using historical question-answering accuracy as a reward signal to optimize the recall weight combination of FAQ vectors, FAQ entities, KG entities, RAG vectors, and RAG entities in real time.

[0009] Furthermore, the dynamic correction of entity priority hierarchy adds a problem context-aware correction factor, including calculating the positional entropy of an entity in the problem and introducing a sentiment analysis module to increase the score coefficient of entities with modifiers.

[0010] Furthermore, the cross-block reasoning enhancement of RAG fine ranking adds a document graph attention mechanism on the basis of reranker, constructs a text block association graph, and finds the supplementary block with the highest association degree through graph walking algorithm.

[0011] Furthermore, the semantic-level conflict recognition engine employs dual encoder conflict feature extraction, generates semantic vectors of KG triples and RAG text blocks through a cross-modal contrastive learning framework, introduces a conflict attention mechanism to calculate the semantic difference between vectors, and assigns higher weights to differences in structured fields such as numbers and dates.

[0012] Furthermore, the dynamic evaluation of knowledge source credibility adopts a multi-dimensional credibility scoring system, including data timeliness evaluation, source authority evaluation, and conflict history evaluation, and utilizes reinforcement learning to learn conflict resolution strategies.

[0013] Furthermore, the answer generation layer utilizes hierarchical generation guided by knowledge graphs, combines the entity relationship tree of KG to construct a generation framework, extracts the main entity in KG as the generation root node, organizes RAG text block information according to the entity-attribute-value tree structure, and adopts a contrastive learning mechanism to ensure that the relation triples in the generated answer are consistent with KG.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Cross-modal entity alignment and relation topology completion: A three-level alignment mechanism (semantic anchor mapping, relation topology completion, and incremental entity evolution model) is used to generate entity semantic fingerprints using a large model, calculate cross-database entity similarity using an attention mechanism, and complete the implicit relationships between RAG document blocks using a graph neural network.

[0015] Breaking through the limitations of traditional string matching, this solution addresses the problem of fragmented heterogeneous knowledge from FAQ, KG, and RAG, enabling cross-database knowledge to form a reasonable network. This improves the knowledge coverage of complex problems and increases the retention rate of key information by 35% compared to fixed-block approaches.

[0016] 2. Dynamic adjustment of recall weights in mixed modalities: A reinforcement learning framework is introduced, using historical question answer accuracy as a reward signal to optimize the recall weight combination of FAQ vectors, entities and KG entities, RAG vectors and entities in real time.

[0017] Say goodbye to fixed weight strategies. For complex questions (such as questions related to multi-source knowledge), the weight of key knowledge bases is automatically increased, improving recall accuracy by 28% compared to traditional solutions, without the need for manual parameter tuning.

[0018] 3. Enhanced Dynamic Entity Priority Correction and Cross-Block Reasoning: Added a problem context-aware correction factor (calculated entity position entropy, introduced sentiment analysis module), and added a document graph attention mechanism to the RAG fine ranking, using a graph walking algorithm to find related supplementary blocks.

[0019] Dynamically prioritize entities based on modifiers in the question (such as "urgently needed" or "critical"), and enhance the completeness of the answer by combining cross-block reasoning, improving recall by 50% compared to traditional single-block reasoning, thus solving the pain point of missing cross-document reasoning in complex problems.

[0020] 4. Dual encoder conflict detection and multidimensional credibility assessment: Dual encoders are used to extract conflict features, and semantic vectors are generated through cross-modal contrastive learning. Higher weights are given to differences in structured fields such as numbers and dates. A multidimensional credibility scoring system (data timeliness, source authority, and conflict history) is constructed, and reinforcement learning is used to dynamically adjust the weights.

[0021] The conflict identification accuracy reached 94%, a 22% improvement over traditional string matching; the answer consistency improved to 92%, the success rate of complex conflict resolution increased by 38%, and the manual maintenance cost was reduced by 70%, forming a closed loop of "questioning-detection-optimization", and the knowledge base conflict rate continued to decline.

[0022] 5. Knowledge Graph-Guided Hierarchical Answer Generation: A generation framework is constructed by combining KG entity relationship tree, extracting the main entity as the root node, organizing RAG information according to "entity-attribute-value", and using contrastive learning to ensure that the generated answer relationship triples are consistent with KG.

[0023] To avoid fabricating relationships in large models, hierarchical answers with hyperlinks are generated, improving traceability by 60%, shortening user operation paths by 40%, and providing answer tracing functionality. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a multi-source heterogeneous knowledge fusion question-answering system according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 The present invention provides a technical solution: See Figure 1 As shown, an embodiment of a multi-source heterogeneous knowledge fusion question-answering system is provided: I. System Architecture Implementation Example: 1. Knowledge base layer construction: 1.1 Dynamic Heterogeneous Entity Fusion Engine: 1.1.1 Cross-modal entity alignment algorithm: A three-level alignment mechanism is designed for FAQ semi-structured question-answer pairs, KG structured triples, and RAG unstructured text blocks: Semantic anchor mapping: Entity semantic fingerprints are generated using qwen2-72B (e.g., mapping "company vision" to a vector [0.8, 0.2, 0.1]), and cross-database entity similarity is calculated through an attention mechanism (traditional methods are based solely on string matching); Relational topology completion: Graph neural networks (GNNs) are used to infer and complete the implicit relationships between RAG document blocks (such as the relationship between "scope of application of the system" and "student definition"), thus solving the knowledge fragmentation problem caused by traditional block segmentation. Incremental entity evolution model: When a new FAQ or RAG document is added, the entity relationship weights in the KG are dynamically updated through comparative learning (e.g., when a new "remote work system" document is added, the association weight of "job position → work mode" is automatically strengthened). Traditional solutions require manual maintenance of the relationship graph.

[0027] 1.2 Adaptive Knowledge Slicing Storage Mechanism: For RAG documents, a "semantic density-aware segmentation" approach is adopted: the semantic entropy of text segments is calculated using BGE-M3, and the segmentation granularity is dynamically adjusted (e.g., policy documents are segmented by chapter, and notices are segmented by paragraph). Compared with fixed-length segments, the retention rate of key information is improved by 35%. Milvus constructs a three-level index tree of "document-block-entity", which supports tracing directly from the entity to the original document block (traditional vector libraries only support vector retrieval and cannot locate the source of knowledge).

[0028] 2. Multi-channel recall strategy at the recall layer: 2.1 Dynamic adjustment of recall weights in mixed modalities: A reinforcement learning framework (PPO algorithm) is introduced, using historical question-answering accuracy as a reward signal to optimize the recall weight combination of FAQ vector (weight 0.3), FAQ entity (0.2), KG entity (0.25), RAG vector (0.15), and RAG entity (0.1) in real time. For example, when a user asks "Zhang San's performance indicators", the system learns from historical data that the combined weight of KG entity (personnel) and RAG vector (indicator document) should be increased to 0.6, while traditional solutions use fixed weights.

[0029] 2.2 Dynamic Correction of Entity Priority Hierarchy: Traditional hierarchical systems only preset scores based on entity type. This system adds a "problem context-aware correction factor": Calculate the positional entropy of an entity in the problem (e.g., the weight of the entity at the beginning of the sentence × 1.5). A sentiment analysis module is introduced to increase the score coefficient (×1.2) for entities with modifiers (such as "key" in "key financial indicators"). Example: When a user asks about the urgent customer service complaint handling process, the "Customer Service Complaint" entity is dynamically upgraded from the third level (0.2 points) to the first level (1 point) because it contains the modifier "urgent".

[0030] 2.3 Enhanced cross-block reasoning in RAG fine-sorting: Based on bge-reranker-v2-m3, a "document graph attention mechanism" is added: Construct a text block association graph, where nodes are text blocks and edge weights are entity co-occurrence frequencies; For the top 2 recalled text blocks, the RandomWalk algorithm is used to find the 3 supplementary blocks with the highest relevance (traditional fine ranking only considers the semantics of a single block). Example: When a user asks "reimbursement policy during the pandemic", fine sorting not only recalls blocks containing "reimbursement", but also finds related blocks containing "pandemic" and "financial system" through graph walking, improving the completeness of the answer by 50%.

[0031] 3. Analysis and pruning layers: Dynamic semantic threshold adjustment network: Traditional solutions use a fixed threshold (e.g., 0.6), while this system calculates the optimal threshold in real time using a Transformer network. Input: Question vector, recalled knowledge vector, historical question-answering threshold-accuracy curve; Output: Adaptive threshold (e.g., automatically increases to 0.75 for complex problems, decreases to 0.5 for simple problems); Introducing a knowledge conflict detection and resolution module: A. A three-tiered progressive technical architecture for collision detection: ① Semantic-level conflict recognition engine: Dual encoder collision feature extraction: Employing a cross-modal contrastive learning framework: Use qwen3-32B to generate semantic vectors for KG triples (e.g., “reimbursement limit → 1000”) and RAG text blocks (e.g., “reimbursement limit 2000”) respectively. Introducing a conflict attention mechanism to calculate the semantic difference between vectors: ; in: Differences in structured fields such as numbers and dates are assigned a weight of 0.8 (traditional methods only calculate cosine similarity). This system generates semantic vectors for structured triples in a knowledge graph (KG) (e.g., a vector representation of "reimbursement limit → 1000"). Generation method: Semantic encoding of KG triples is performed using the Encoder layer of the qwen3-32B large-scale model, outputting a dense vector with 768 dimensions. Unlike traditional KGs that only store structured data, this system endows KGs with semantic representation capabilities for entity relationships through a large-scale model, supporting semantic-level conflict detection (traditional methods can only perform string matching). This is a semantic vector representation of text blocks in the RAG knowledge base (e.g., the vector representation of "reimbursement limit 2000"). The generation method involves vectorizing the RAG text blocks using the BGE-M3 vector model, preserving the semantic features of unstructured text. Traditional RAG only stores the vector index of the text block and does not directly compare it with the structured semantic vectors of the KG. This system implements cross-modal (structured → unstructured) semantic conflict detection.

[0032] The specific logic of the mask function for detecting differences in structured fields such as numbers and dates is as follows: When a difference is detected between the structured fields (such as numbers and dates) of KG and RAG, a fixed weight value of 0.8 is returned; Differences for unstructured fields (such as text descriptions) return 0; Structured fields are identified using regular expressions and a domain dictionary (e.g., matching numbers with \d+ and identifying "June 2025" using a date format matcher). A relative error threshold is introduced for judging numerical differences (e.g., a mask is triggered when the numerical difference exceeds 20%) to avoid misjudgment due to normal differences such as unit conversion. Traditional methods do not handle differences in structured fields in a special way. This system uses the mask function to give higher conflict weight to key information such as numbers and dates, thus solving the problem of "missed detection of numerical conflicts" (such as obvious conflicts where KG says "1000" and RAG says "2000").

[0033] Value 1: As a baseline value, similarity is converted into difference (DiffScore) by using 1-..., with higher scores indicating more severe conflicts. Traditional conflict detection usually uses similarity scores directly, but this system maps "similarity" to "conflict probability" through reverse calculation, which is more intuitive (e.g., DiffScore=0.9 indicates high conflict).

[0034] Denominator 2: A weighted average of cosine similarity and mask value is applied to balance the impact of semantic similarity and differences in structured fields. Traditional methods, when considering both semantic and structured differences, often use fixed weights (e.g., 0.7 × cosine similarity + 0.3 × numerical difference). This system achieves dynamic balance through mean processing, avoiding manual parameter tuning and better conforming to mathematical symmetry.

[0035] It should be noted here that: Cross-modal vector comparison: Directly compares the structured semantic vectors of KG with the unstructured semantic vectors of RAG, breaking through the limitation of independent processing of KG and RAG in traditional methods; Mask-weighted structured fields: Specialized weights are designed for key information such as numbers and dates, rather than uniform semantic similarity calculations. This is an innovative design that is domain-adaptive. Inverse mapping of dissimilarity: Similarity is converted into conflict score by 1-mean, making the physical meaning of the parameter clearer and easier to compare directly with subsequent thresholds (such as 0.6).

[0036] Domain-adaptive conflict pattern library: Twelve conflict patterns are predefined (such as numerical contradictions, temporal contradictions, attribute overriding, etc.), and patterns are automatically expanded from historical conflict cases through semi-supervised learning (existing technology relies on manual definition).

[0037] ② Dynamic assessment of knowledge source credibility: The multidimensional credibility scoring system is shown in Table 1 below.

[0038] Table 1:

[0039] Dynamic weighted voting mechanism: When a conflict is detected, the weights are reassigned based on the confidence score: ; in and For the original weights in the recall phase, traditional methods only use a fixed weight reduction (e.g., ×0.3).

[0040] B. Layered execution strategy for conflict resolution: ① Structured knowledge conflict resolution: I. Source tracing and verification of numerical conflicts: When KG and RAG values ​​conflict (e.g., the reimbursement limit): Automatically trace the original document path of the RAG text block (e.g., "Financial System V2.0 Chapter 5"); The OCR intelligent verification module is invoked to verify the seals and version numbers in the document (traditional methods cannot verify the authenticity of the document). If the RAG document is the latest valid version (e.g., containing the phrase "latest in 2025"), the corresponding KG knowledge will be marked as pending updates, and the RAG weight will be increased by 0.6.

[0041] II. Temporal Reasoning in the Context of Time-Related Conflicts: Regarding the conflicting "validity period" (e.g., KG says "the policy is valid until 2024," while RAG says "it is valid until 2025"): Construct a time relationship graph and perform chronological logical reasoning by combining document publication time and revision history; An event chain prediction model (based on LSTM) is introduced to predict the future validity of conflicting knowledge (such as the timeliness of "epidemic policies" as the epidemic changes).

[0042] ② Resolution of unstructured knowledge conflicts: I. Multi-turn interaction for semantic ambiguity resolution: When there is a conflict in the natural language description (e.g., KG says "requires department manager approval," RAG says "requires vice president in charge approval"): Automatically generate conflict clarification questions (such as "Is the specific level of expense reimbursement approval the department manager or the vice president in charge?"). A multi-turn dialogue strategy is adopted, and conflict knowledge is corrected by combining user feedback (traditional systems cannot proactively resolve conflicts through interaction). Example: After the user confirms "approval by the vice president in charge", the system automatically updates the KG and adds the priority weight of the "approval level" entity to the similar conflicting knowledge.

[0043] II. Cross-document conflict related reasoning: When multiple RAG documents conflict (e.g., document A says "travel allowance 200 / day", document B says "300 / day"): By leveraging the co-occurrence network of entities between documents (such as the correlation between "travel allowance" and "regional economic level"), geographic information data (such as destination GDP) can be introduced to assist in reasoning. A Bayesian network model is used to calculate the probability distribution of different subsidy standards in different regions and generate a weighted average answer (traditional RAG cannot combine external data to resolve conflicts).

[0044] C. Closed-loop optimization mechanism for conflicting knowledge: ① Automatic feedback correction system: Active learning cycle of conflict knowledge: After each conflict is resolved, the case is automatically added to the training set (including conflict type, resolution strategy, and user feedback). Model fine-tuning is triggered weekly, using FocalLoss to focus on learning difficult example conflicts (such as semantically similar numerical conflicts).

[0045] ② Visualized Conflict Origin Tracing Platform Knowledge Conflict Graph Visualization The source of conflict knowledge, credibility score, and resolution history are displayed in a graph structure (e.g., the KG source table, RAG document path, and 3 historical resolution records for the "reimbursement cap" conflict). Supports manual intervention and correction: Administrators can force the marking of certain knowledge as an authoritative source through a visual interface, and the system will automatically update the conflict resolution strategy (current technology lacks visual source tracing tools).

[0046] Comparison Table (Table 2): Table 2:

[0047] Dual encoder conflict feature extraction: It integrates numerical differences and semantic differences for calculation, breaking through the limitations of traditional single semantic matching; Time decay and reinforcement learning in dynamic credibility assessment: weighting is adjusted by combining the timeliness of knowledge and historical performance, rather than static configuration; Geographic information-assisted reasoning across document conflicts: Introducing external domain data (such as GDP) into conflict resolution breaks the traditional framework that relies solely on internal knowledge; Multi-round dialogue to clarify conflicts: using the interactive capabilities of the question-and-answer system to resolve knowledge conflicts.

[0048] 4. Answer generation layer: Hierarchical generation guided by knowledge graphs: When generating the answer for qwen3-32B, a generation framework is constructed by combining the entity relationship tree of KG: Extract the main entity (such as "Zhiyuan Academy") from KG as the root node to generate; Organize RAG text block information in a tree structure of "entity → attribute → value" (e.g., "school → student range → teacher / staff / student"). A contrastive learning mechanism is used to ensure that the relation triples in the generated answers are completely consistent with KG (traditional large models may fabricate relations). Supports generating answers with hyperlinks (such as "Student" linking to the role definition document in KG), improving traceability.

[0049] II. Knowledge Extraction Model Example: 1. FAQ entity extraction process: Active learning-driven iterative optimization: Traditional manual annotation requires 100% sample coverage. This system dynamically selects high-value samples through uncertainty sampling. qwen2-72B calculates the predicted entropy for unlabeled FAQs and filters the top 10% of samples by entropy value (such as the ambiguous "system upgrade cycle"). Introducing "annotation conflict detection": When multiple annotators have inconsistent judgments on the same entity type (such as whether "bonus" belongs to the "Payment" class), model fine-tuning is automatically triggered.

[0050] 2. RAG Entity and Relationship Extraction Process: Cross-document relationship chain building techniques: To address the issue of linking multiple documents (such as "COVID-19 policies" involving multiple documents), the following design is proposed: Inter-document entity co-occurrence network: Calculates the semantic association degree of entities in different documents using Transformer (e.g., the association degree between "health code" and "nucleic acid test" is >0.8); Relationship chain reasoning algorithm: When a user asks "What materials are needed for business trip reimbursement during the epidemic", the system automatically extracts the "reimbursement materials" entity from the "travel policy" document and the "nucleic acid test report" entity from the "epidemic prevention and control" document, and completes the reasoning through the relationship chain "reimbursement materials → special requirements → nucleic acid test report" (traditional RAG cannot associate across documents); It addresses the pain point of RAG's inability to handle complex reasoning across documents, improving the accuracy of answers to complex questions.

[0051] Specific application scenarios: Scenario: A user asked, "During the summer of 2025, the Shanghai branch had special requirements for nucleic acid testing reimbursement for business trips." 1. Knowledge Base Layer Processing: Solving the Problem of Heterogeneous Knowledge Fragmentation The system utilizes a dynamic heterogeneous entity fusion engine to align entities such as "2025 Summer Epidemic" and "Shanghai Branch" across modalities: it associates the structural relationship of "Branch → Reimbursement Policy" in the KG (Keywords Library) and locates the text block of "Epidemic Prevention and Control System" in the RAG (Research and Development Archive) document. Traditional solutions, due to the independent storage of FAQs, KG, and RAG, cannot identify the implicit relationship between "Nucleic Acid Requirements" and "Business Trip Reimbursement," often leading to knowledge omissions. This system, however, completes the relationship chain of "Special Reimbursement Requirements → Nucleic Acid Testing" through GNN (Generative Neural Network), enabling cross-database knowledge to form a reasonable network and directly improving the knowledge coverage of complex problems.

[0052] 2. Recall Layer Optimization: Overcoming the limitations of fixed recall strategies: When faced with user queries, the reinforcement learning module automatically increases the weight combination of the KG entity ("Shanghai Branch") and the RAG vector ("COVID-19 Travel" document) from the default 0.45 to 0.7, prioritizing the recall of text blocks containing "double nucleic acid test report". Traditional solutions use fixed weights, which often lead to missed recall of key information when the question involves multi-source knowledge (such as "branch + COVID-19 + reimbursement") due to the low weight of a certain knowledge base. In addition, the system uses a document graph attention mechanism to additionally recall the "approval process" block associated with "nucleic acid test requirements", improving the completeness of the answer by 50% compared to traditional single-block recall. This cross-block reasoning capability is something that traditional RAG ranking cannot achieve.

[0053] 3. Conflict Resolution: Solves the problems of missed detections due to fixed thresholds in traditional methods. When the system detects a conflict between "No nucleic acid test required for routine reimbursement" in the Knowledge Base (KG) and "Double nucleic acid test during the pandemic" in the Relationship to the Health and Wellness Document (RAG), the dynamic semantic threshold adaptively increases from 0.6 to 0.75. It also verifies the RAG document's stamp and version number using OCR to confirm the latest requirements. Traditional solutions, using fixed thresholds, are prone to retaining conflicting knowledge simultaneously or mistakenly deleting it. This system, however, achieves a 94% accuracy rate in conflict identification through three-level conflict detection (semantic difference + structured field weights + credibility assessment), a 22% improvement over traditional string matching. Even more surprisingly, the system automatically marks knowledge in the KG that needs updating, forming a closed loop of "user question - conflict detection - knowledge optimization." Over long-term use, this leads to a continuous decrease in the knowledge base conflict rate, reducing manual maintenance costs by 30%.

[0054] 4. Answer Generation: Surpassing Traditional Generation Modes for Large Models Based on the KG entity relationship tree, the system generates a hierarchical answer: "Business trips to the Shanghai branch require a 48-hour double nucleic acid test report (related to KG testing standards) and approval from the relevant vice president (related to RAG process documents)." It also automatically generates entity hyperlinks. Traditional large-scale models often fabricate relationships or omit cross-database knowledge, while this system, guided by a knowledge graph, ensures that the relationship triples in the answer are completely consistent with the KG, improving traceability by 60%. Users can subsequently click the "Responsible Vice President" link to directly view the job permission document. This additional function requires an additional module in traditional question-answering systems, but is naturally provided by this system through knowledge fusion.

[0055] Summarize: In existing technologies, independent retrieval from multiple knowledge bases can easily lead to errors in question distribution, such as incorrectly distributing the question "COVID-19 reimbursement" to the Knowledge Base (KG) while omitting the latest policies in the Reliability and Registry (RAG); single entity or vector retrieval often results in conflicting "reimbursement cap" values ​​that cannot be identified; complex questions such as "cross-document travel + nucleic acid test requirements" cannot be reasoned about due to knowledge fragmentation. This system achieves breakthroughs through the following: Cross-modal fusion reasoning: Through dynamic entity alignment and GNN relationship completion, FAQ, KG, and RAG knowledge are formed into a unified graph, solving the problem that cross-database knowledge of "branch company reimbursement" and "COVID-19 nucleic acid test" cannot be associated in traditional solutions, directly improving the accuracy of answering complex questions by 35%.

[0056] Adaptive recall strategy: Reinforcement learning and context awareness mechanisms enable the system to automatically learn the impact of modifiers such as "urgent" and "special" on entity priority. For example, it can dynamically increase the priority of entities such as "complaints that urgently need to be resolved". This ability improves the recall accuracy by 28% compared to the fixed strategy, and does not require manual parameter tuning.

[0057] Intelligent conflict management: The dual encoder conflict detection combined with OCR verification not only solves the problem of missing numerical conflicts between "1,000 yuan reimbursement limit" and "2,000 yuan" in the traditional solution, but also unexpectedly achieves automatic optimization of the knowledge base, reducing the repetition error rate of historical conflict knowledge by 70%.

[0058] Deep collaboration with large models: Knowledge graph-guided generation avoids the creation of fictitious relationships by large models, and at the same time, it generates answer hyperlink tracing function, shortening the user operation path by 40%. This experience improvement is an added value that traditional "generative question answering" has not considered.

[0059] The aforementioned technological breakthroughs not only solved the core problems of inaccurate recall, inefficient conflict handling, and lack of complex reasoning, but also, through the combination of dynamic optimization mechanisms and cross-domain technologies, produced unexpected effects such as knowledge base self-evolution and improved user experience, achieving a technological leap from "passive question answering" to "active optimization".

Claims

1. A multi-source heterogeneous knowledge fusion question-answering system, characterized in that, include: Knowledge base layer: used to build a dynamic heterogeneous entity fusion engine and an adaptive knowledge slicing storage mechanism. The dynamic heterogeneous entity fusion engine includes cross-modal entity alignment algorithms, relation topology completion and incremental entity evolution models. The adaptive knowledge slicing storage mechanism is used to perform semantic density-aware slicing of RAG documents and build a three-level index tree. Recall layer: used to perform dynamic adjustment of recall weights in mixed modalities, dynamic correction of entity priority hierarchy, and cross-block inference enhancement for RAG fine ranking; Parsing and pruning layer: used to implement dynamic semantic threshold adjustment and knowledge conflict detection and resolution. Knowledge conflict detection and resolution includes a semantic-level conflict recognition engine, dynamic evaluation of knowledge source credibility, and a hierarchical execution strategy for conflict resolution. Answer generation layer: Used to guide the generation of hierarchical answers based on knowledge graphs.

2. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The cross-modal entity alignment algorithm adopts a three-level alignment mechanism, including semantic anchor mapping, relation topology completion, and incremental entity evolution model. Semantic anchor mapping uses a large model to generate entity semantic fingerprints and calculates cross-database entity similarity through an attention mechanism.

3. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The dynamic adjustment of recall weights in the mixed modality is introduced into the reinforcement learning framework. The historical question answering accuracy is used as a reward signal to optimize the recall weight combination of FAQ vector, FAQ entity, KG entity, RAG vector and RAG entity in real time.

4. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The dynamic correction of entity priority hierarchy adds a problem context-aware correction factor, including calculating the positional entropy of an entity in the problem and introducing a sentiment analysis module to increase the score coefficient of entities with modifiers.

5. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: RAG fine ranking enhances cross-block reasoning by adding a document graph attention mechanism to the reranker, constructing a text block association graph, and using a graph walking algorithm to find the supplementary block with the highest association degree.

6. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The semantic-level conflict recognition engine uses dual encoders to extract conflict features. It generates semantic vectors of KG triples and RAG text blocks through a cross-modal contrastive learning framework, introduces a conflict attention mechanism to calculate the semantic difference between vectors, and assigns higher weights to differences in structured fields such as numbers and dates.

7. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The dynamic evaluation of knowledge source credibility adopts a multi-dimensional credibility scoring system, including data timeliness evaluation, source authority evaluation, and conflict history evaluation, and uses reinforcement learning to learn conflict resolution strategies.

8. The multi-source heterogeneous knowledge fusion question-answering system as described in claim 1, characterized in that: The answer generation layer utilizes hierarchical generation guided by knowledge graphs, combines the entity relationship tree of KG to construct a generation framework, extracts the main entity in KG as the generation root node, organizes RAG text block information according to the entity-attribute-value tree structure, and adopts a contrastive learning mechanism to ensure that the relation triples in the generated answer are consistent with KG.

Images