A dual-channel retrieval-guided evidence graph reasoning enhanced generation method

Abstract

This invention proposes a dual-channel retrieval-guided method for enhancing evidence graph reasoning generation, belonging to the fields of knowledge graph and natural language processing. It receives and parses heterogeneous domain knowledge graphs, constructing a pattern layer representation. An adaptive dual-channel collaborative retrieval method is designed, performing weighted fusion and consistency / cross-validation on the retrieval results from both channels to obtain a set of high-confidence evidence seed nodes and form conflict markers. Based on this, evidence graph expansion and pruning are performed to obtain compact, high-quality evidence subgraphs. Graph neural network reasoning and multi-hop reasoning path discovery are carried out on the evidence subgraphs, and the evidence subgraphs and reasoning paths are represented in a structured, sequential manner. The generated results undergo multiple verifications and confidence assessments. This invention improves interpretability through dual-channel collaborative retrieval and evidence graph organization of reasoning basis, and enhances the accuracy and reliability of the generated results through multiple verifications and hierarchical output mechanisms. It is applicable to scenarios such as intelligent question answering, knowledge reasoning, and decision support.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph and natural language processing technology, specifically, it relates to a dual-channel retrieval-guided evidence graph reasoning enhancement generation method. Background Technology

[0002] With the rapid development of artificial intelligence technology, large-scale language models have demonstrated powerful capabilities in natural language understanding and generation tasks. Representative works such as OpenAI's GPT series models, Google's PaLM model, Meta's LLaMA model, and Anthropic's Claude model have achieved groundbreaking results in multiple natural language processing benchmarks. However, these models still face challenges in handling knowledge-intensive tasks, including lagging knowledge updates, susceptibility to illusions, and opaque reasoning processes. To address these issues, Retrieval-Augmented Generation (RAG) technology has emerged. The RAG model proposed by Lewis et al. in 2020 is a pioneering work in this field. This method combines a pre-trained dense retrieval engine with a seq2seq generative model, achieving significant improvements in open-domain question answering tasks. This technique effectively enhances the model's performance in tasks such as question answering, information extraction, and content generation by retrieving relevant information from external knowledge bases and providing it as context to the language model. In recent years, with the development of technology, improved methods such as Self-RAG, CRAG (Corrective RAG), and Adaptive RAG have emerged. These works have optimized the RAG framework from the perspectives of retrieval timing, retrieval quality assessment, and adaptive retrieval strategies.

[0003] Traditional retrieval augmentation systems primarily employ vector similarity-based retrieval methods. Specifically, the system first uses a pre-trained text embedding model to convert user queries and knowledge base documents into vector representations. Commonly used embedding models include BERT, Sentence-BERT, OpenAI's text-embedding series, and Cohere's embedding model. Then, it retrieves the most relevant document fragments by calculating the similarity between vectors. While this method can provide relevant information to some extent, it has significant limitations. First, the flattened data representation based on vectors struggles to capture the complex semantic relationships and logical dependencies between entities, resulting in fragmented information fragments lacking inherent coherence in the retrieval results. Second, for complex queries involving multiple entity associations, traditional methods often only retrieve content related to each keyword separately, failing to integrate them into a coherent reasoning chain. Third, due to the diversity and ambiguity of natural language expressions, vector similarity-based retrieval lacks accuracy when handling domain-specific knowledge and is prone to introducing noise. These problems severely impact the reliability and usability of the system in practical applications.

[0004] To overcome the shortcomings of traditional methods, researchers have begun exploring the integration of knowledge graphs into retrieval-enhanced generative systems. Knowledge graphs organize knowledge in a graph structure, representing entities through nodes and relationships between entities through edges, naturally depicting complex relationships in the real world. Compared to flat text fragments, knowledge graphs provide structured knowledge representations, enabling systems to perform multi-hop reasoning along paths in the graph, thus better understanding and answering complex questions. Current research attempts to integrate knowledge graphs into retrieval-enhanced generative processes, such as retrieving relevant triples from knowledge graphs and adding them as text input to language models. Sun et al.'s GREASELM method combines a language model with a knowledge graph encoder, encoding the retrieved subgraphs through graph neural networks; Baek et al.'s KAPING method explores knowledge-aware path selection strategies; and Luo et al.'s RoG (Reasoning on Graphs) method attempts to enable large language models to plan and reason on knowledge graphs. However, these methods primarily focus on graph reasoning itself, often using simple entity links or keyword matching for initial retrieval, failing to fully consider retrieval needs at different granularities. However, this simple integration method does not fully utilize the graph structure characteristics of knowledge graphs. It merely linearizes structured knowledge into text sequences. For example, the KG-to-Text method converts triples into natural language descriptions, losing the topological information of relationships between entities, and is difficult to support complex reasoning needs.

[0005] From a query understanding perspective, user queries can be broadly categorized into two types: detailed queries targeting specific entities or relationships, which typically involve specific factual information and require precise matching of relevant entities; and global queries involving abstract concepts or topics, which focus on broader semantic categories and conceptual connections. In the field of information retrieval, this distinction is similar to Broder's classification of query intent: navigational, informational, and transactional queries. Different types of queries require different retrieval strategies. Traditional single retrieval strategies struggle to simultaneously meet the needs of both types of queries. For detailed queries, the system needs to quickly locate specific entity nodes and their neighboring relationships; while for global queries, the system needs to understand the query intent at a higher semantic level and aggregate relevant conceptual and topical information. Existing work, such as HybridQA, attempts to combine tables and text for hybrid retrieval, but lacks in-depth utilization of the structural characteristics of knowledge graphs. Therefore, designing a retrieval mechanism that can adaptively handle query needs at different levels becomes crucial for improving system performance.

[0006] In utilizing knowledge graphs, existing methods often lack effective organization and representation when transforming retrieved graph information into evidence usable for reasoning. Traditional subgraph extraction methods, such as Personalized PageRank (PPR) and k-hop neighbor sampling, are mainly based on graph topology and fail to fully consider semantic relevance to the query. Simply listing retrieved triples for a language model leads to information overload and reasoning confusion. In reality, for a given query, there are multiple possible reasoning paths in the knowledge graph, each containing different chains of evidence. Xiong et al. proposed a path ranking method in their research on multi-hop reasoning, but this method is mainly for simple chain-like reasoning and struggles to handle complex graph-like reasoning structures. How to extract the most query-relevant subgraph structures from large-scale knowledge graphs and organize them into evidence graphs conducive to reasoning is a pressing problem. An ideal evidence graph should contain both local details and global association patterns, providing sufficient yet non-redundant knowledge support for subsequent reasoning.

[0007] During the generation phase, even with relevant retrieval results, language models may still encounter factual errors or logical inconsistencies when generating answers. This is because the generation process of language models is essentially a sequence prediction based on probability distributions, lacking explicit constraints on knowledge consistency. Some existing studies attempt to improve generation quality by allowing the model to self-reflect or validate, such as using a chain-of-thought approach to guide the model through step-by-step reasoning, or generating multiple times and selecting the optimal result. The Chain-of-Thought hint method proposed by Wei et al. significantly improves the performance of complex reasoning tasks by guiding the model to gradually reveal the reasoning process; the Tree-of-Thoughts method by Yao et al. organizes the reasoning process into a tree structure, supporting backtracking and exploration; and the Self-Consistency method proposed by Wang et al. improves the reliability of answers by sampling multiple reasoning paths and voting. However, these methods mainly rely on the model's own capabilities, lacking external knowledge constraints and validation, and have limited effectiveness when dealing with tasks requiring strict factual accuracy. The RARR (Retrofitting RAG with Retrieval) method proposed by Gao et al. attempts to perform fact verification and correction after generation, but its verification mechanism is relatively simple, mainly relying on re-retrieval and similarity comparison. An ideal verification mechanism should combine multiple strategies: on the one hand, verify the coherence and rationality of the logical chain from the perspective of the reasoning process; on the other hand, examine whether the generated result conforms to the facts and rules in the knowledge graph from the perspective of knowledge constraints.

[0008] Furthermore, in real-world applications, user input often goes beyond plain text, potentially including multimodal information such as images, tables, and charts. For example, in intelligent question-answering systems, a user might upload an image containing complex information and ask a question; in form-filling scenarios, the form itself might contain fields and instructions in various formats. The emergence of large-scale multimodal language models has made it possible to handle such complex inputs. Representative multimodal models include OpenAI's GPT-4V, Google's Gemini, and Anthropic's Claude 3 series. These models, by fusing visual encoders and language models, can understand information from multiple modalities, including images and text. However, most existing retrieval-enhanced generation systems are designed only for text input and lack the ability to process multimodal inputs, limiting their application in real-world scenarios. While a few works, such as RA-CM3, attempt to extend retrieval enhancement to multimodal scenarios, these methods primarily focus on image-text retrieval and fail to effectively integrate the structured knowledge of knowledge graphs. How to effectively combine multimodal understanding capabilities with knowledge graph retrieval and reasoning to build an end-to-end multimodal intelligent generation system is a significant challenge in current research.

[0009] In summary, existing retrieval-enhanced generation technologies have shortcomings in areas such as knowledge graph utilization, retrieval strategy design, reasoning process organization, result verification, and multimodal input processing. Specifically, these shortcomings manifest as follows: a single retrieval level, failing to adapt to information needs at different granularities; insufficient utilization of graph structure information and a lack of effective evidence graph construction mechanisms; a lack of interpretability in the reasoning process, making it difficult to guarantee the reliability of generated results; an imperfect verification mechanism, unable to comprehensively utilize multiple strategies for verification; and insufficient multimodal understanding capabilities, limiting applicability to specific scenarios. Therefore, there is an urgent need for an innovative technical solution that can systematically address these issues, developing an intelligent generation technology that integrates dual-channel retrieval, evidence graph reasoning, multi-dimensional verification, and multimodal understanding to meet the practical application requirements for accuracy, reliability, and flexibility. Summary of the Invention

[0010] To address the problems of insufficient retrieval accuracy, opaque reasoning paths, and lack of verifiability of generated results in existing knowledge augmentation generation technologies, this invention proposes a dual-channel retrieval-guided evidence graph reasoning augmentation generation method. By integrating a two-layer retrieval mechanism, evidence graph construction technology, and multiple verification strategies, this method can achieve high-precision knowledge retrieval, an interpretable reasoning process, and reliable generated results, and is widely applicable to various application scenarios such as intelligent question answering and knowledge reasoning.

[0011] This invention is achieved through the following technical solution: a dual-channel retrieval-guided method for enhancing the generation of evidence graph reasoning. Step 1: Receive and parse the multimodal domain heterogeneous knowledge graph, construct the pattern layer representation; perform multi-granularity embedding representation learning at the entity level, relation level and path level respectively, and establish a hybrid indexing system and mapping synchronization mechanism of graph database and vector database; Step 2, design adaptive dual-channel collaborative retrieval: perform intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering, while the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database; Step 3: Perform weighted fusion and consistency / cross-validation on the retrieval results of the two channels to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, expand and prune the evidence graph to obtain a compact and high-quality evidence subgraph. Step 4: Perform graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serialize and encode candidate paths and calculate path confidence, and screen Top-M high-quality reasoning paths; transform the reasoning paths into structured evidence chains and perform confidence propagation and aggregation. Step 5: The evidence subgraph and reasoning path are represented in a structured and sequential manner. A unified prompt word framework containing "query restatement - evidence presentation - generation instruction" is constructed to guide the large language model to generate the answer and record the evidence identifiers of key arguments to achieve answer-evidence alignment. When the generation quality is not up to standard, a supplementary retrieval is triggered and the process of evidence construction and reasoning is returned to form a closed loop. Step 6: Perform multiple verifications and confidence assessments on the generated results, including direct verification using the knowledge graph, verification through evidence chain reasoning, dual-channel consistency testing, verification of redundant evidence, and conflict detection and resolution. Based on weighted fusion, obtain a comprehensive confidence level to determine whether to output directly, label uncertainties, or trigger closed-loop error correction.

[0012] Furthermore, step 1 specifically includes: Step 1.1: Heterogeneous knowledge graph reception and parsing. Receive complete domain knowledge graph data, including entity sets, relation sets, attribute sets, and their corresponding metadata information; The knowledge graph is structured and parsed to identify meta-information including hierarchical relationships of entity types, semantic categories of relationship types, and attribute constraint rules, and to construct a schema layer representation of the knowledge graph. Step 1.2: Multi-granularity graph embedding representation learning, adopting a multi-level graph representation learning strategy to capture both local structural features and global semantic information; Step 1.3: Construct a hybrid indexing system; For the graph database index layer, Neo4j is used to build native graph storage, node attribute indexes and relation type indexes are created, and composite indexes are created to support multi-condition filtering queries. For the vector database index layer, entity embedding, relation embedding, and path embedding are stored in different sets of the Milvus vector database. A hierarchical navigation small world graph (HNSW) index is built for each set, and a multi-level index structure is set to balance retrieval speed and recall. Establish a mapping table between the graph database and the vector database, and associate the structured representation and vector representation of the same entity through a unique identifier. At the same time, design an index synchronization update strategy to update the data of the two index layers synchronously when the knowledge graph is incrementally updated.

[0013] Furthermore, step 1.2 specifically includes: Step 1.2.1: Entity-level embedding generation. The domain-adaptive language model BGE-Embedding is used to encode entity names, aliases, and descriptive text to generate context-aware semantic vectors. A contrastive learning mechanism is introduced to enhance the discriminativeness of entity embeddings through comparative training of synonymous entity pairs and heteronymous entity pairs. Step 1.2.2: Relationship-aware structured embedding. The Rotat rotating knowledge graph embedding method is used to learn the representation of entities in the relation space. CompGCN is used to fuse relation type information. The features of neighbor nodes and relations are aggregated through message passing mechanism. Weight matrices are assigned to different relation types to capture the semantic differences of heterogeneous relations. Step 1.2.3: Path-level semantic embedding, identify meta-path patterns in the knowledge graph, use TransformerEncoder to sequence encode the meta-paths, learn the vector representation of multi-hop reasoning patterns, and build a path pattern library.

[0014] Furthermore, step 2 specifically includes: Step 2.1: Understanding and decomposing the query intent; Step 2.1.1: Query type identification. The user query intent is classified using a large language model based on Prompt Engineering, which is divided into factual queries, inference queries, multi-hop relational queries, and aggregate statistical queries. The structured information in the query is extracted, including key entities, relational constraints, attribute filtering conditions, and time limits. Step 2.1.2: Query expansion and rewriting. The query is expanded using the domain ontology to add semantically related candidate words. Qwen3 is used to rewrite the query to generate multiple semantically equivalent but different query variants. Disambiguation is performed on fuzzy queries. Step 2.2: Semantic retrieval channel design; Step 2.2.1: Multi-granular semantic matching. The query is encoded in a fine-grained manner, and semantic representations at the word, phrase, and sentence levels are extracted respectively. Multi-granular retrieval is performed in the vector database. For entity-level retrieval, the Top-K1 entity nodes that are most similar to the query semantics are recalled; for relation-level retrieval, the Top-K2 relation types that match the query relation pattern are recalled; for path-level retrieval, the Top-K3 meta-path patterns that are similar to the query inference link are recalled. Step 2.2.2: Re-ranking after retrieval. The ColBERTv2 interaction model is used to refine and re-rank the candidate results. The maximum similarity score between each token in the query and the token in the candidate entity description is calculated, and the scores are accumulated to obtain the overall relevance score. Step 2.3: Design of structured retrieval channels; Step 2.3.1: Query statement generation. The natural language query is converted into Cypher graph query language, enabling the large language model to generate structured query statements, and the generated query statements are subjected to syntax validation and logical consistency checks. Step 2.3.2: Constraint filtering, performing exact match queries with multiple constraints in the graph database; Step 2.3.3: Extract local subgraphs from the matching results, including nodes that satisfy the constraints and their neighboring nodes, and retain the topological structure information and edge weight information of the subgraph; Step 2.4: Dual-channel fusion and evidence node screening; Step 2.4.1: Design the fusion strategy. A weighted fusion method is adopted to dynamically adjust the weight coefficients of the two channels according to the query type. For factual queries, the weight of the structured channel is increased; for open reasoning queries, the weight of the semantic channel is increased. The optimal fusion weight is learned using a ranking learning model. Step 2.4.2: Consistency verification and cross-validation. Calculate the intersection nodes of the retrieval results of the two channels as high-confidence evidence seeds. Perform cross-validation on nodes that appear only in one channel, check their relevance scores in the other channel, mark conflicting nodes, and provide uncertainty information for subsequent reasoning. Step 2.4.3: Initial screening and sorting of evidence nodes. Based on semantic similarity, structural matching degree and node importance, candidate nodes are scored. The scores are processed using softmax normalization to generate the confidence distribution of evidence nodes. The top-N high-confidence nodes are selected as the seed node set for the construction of the evidence graph.

[0015] Furthermore, step 3 specifically includes: Step 3.1: Using the evidence nodes selected in Step 2 as the seed node set, assign an initial activation value to each seed node. The activation value is proportional to the confidence level of the node. Step 3.2: Design a multi-strategy graph expansion mechanism to expand the initial seed node; Step 3.2.1: Attention-based neighbor sampling. For each seed node, the importance weight of its neighbor nodes is calculated using a graph attention mechanism. The sampling probability of different types of neighbors is adjusted according to the semantic relevance of the relation type. A Top-K sampling strategy is adopted to select the K most relevant neighbors for each node for expansion. Step 3.2.2: Meta-path-guided expansion. Using the meta-path pattern library built in Step 1, identify the inference path templates applicable to the query and expand in a targeted manner along the predefined meta-paths. Step 3.2.3: Reinforcement learning-driven exploration strategy, modeling the graph expansion process as a Markov decision process, and using a deep Q-network to learn the optimal expansion strategy; Step 3.3: Refine and prune the evidence subgraph to generate the evidence graph; Step 3.3.1: Node importance assessment. GraphSAINT, a graph neural network, is used to learn the representation of the expanded subgraph. The centrality index of each node is calculated and the PageRank algorithm is used to evaluate the global importance of the node in the subgraph and its relevance to the query. Step 3.3.2: Path quality scoring, taking into account path length, semantic coherence of relationships on the path, and relevance of the endpoint node, to score all paths starting from the seed node; Step 3.3.3: Identify and remove redundant nodes and edges that are irrelevant to the query. Use the Louvain community detection algorithm to identify the module structure in the subgraph, retain the core communities that are relevant to the query, set a relevance threshold, and prune nodes and edges with scores below the threshold. Step 3.3.4: For complex queries, generate multiple evidence subgraphs from different perspectives.

[0016] Furthermore, step 4 specifically includes: Step 4.1: Graph Neural Network Inference; Step 4.1.1: Model the evidence subgraph using R-GCN, assign independent embedding spaces and transformation matrices to different types of nodes and edges, aggregate neighbor information through multi-layer message passing, and update the hidden state of nodes; Step 4.1.2: Introduce cross-layer connections and attention mechanisms in the multi-layer GNN so that the node representation of each layer can receive information from other layers, capture structural features at different scales, and use attention weights to dynamically fuse the representations of different layers to generate the final node embedding. Step 4.1.3: Inject the query vector as additional input into each layer of the GNN, use a gating mechanism to adjust the influence of the query information on the node representation, and learn the subgraph representation that is highly related to the query; Step 4.2: Multi-hop reasoning path discovery, identifying multiple reasoning paths from the query entity to the target entity in the evidence subgraph; Step 4.2.1: Enumerate all feasible paths from the query-related entity to the potential answer entity in the evidence subgraph, and use depth-first search combined with heuristic pruning strategy to enumerate candidate paths; Step 4.2.2: Represent each path as an alternating sequence of entity-relationships and encode the path sequence, introducing position encoding and relation type embedding to enhance the semantic expressiveness of the path and realize path semantic modeling; Step 4.2.3: Calculate the confidence score for each path, and comprehensively consider path length penalty, semantic coherence of relations, and entity relevance. Use a ranking learning model to rank the paths and retain the Top-M high-quality inference paths. Step 4.3: Generate the inference evidence chain, transform the high-confidence inference path into a structured evidence chain, and label the relationship type and confidence score of each edge; Step 4.3.1: Transform the filtered reasoning paths into a structured chain of evidence, label each edge with the relationship type, relationship strength, and number of supporting evidence, construct the hierarchical structure of the chain of evidence, and distinguish between the main reasoning path and the auxiliary supporting path; Step 4.3.2: Use a probabilistic graphical model to model the confidence of the evidence chain. Starting from the seed node, propagate the confidence along the path, considering the reliability of the edges and the prior confidence of the nodes, and aggregate the confidence of multiple evidence chains.

[0017] Furthermore, step 5 specifically includes: Step 5.1: Structured Representation of Evidence Graphs; Transform the evidence subgraphs and reasoning paths constructed in Steps 3 and 4 into structured text representations to achieve evidence subgraph serialization; Design a unified prompt word framework, which includes three core parts: query restatement, evidence presentation, and generation instructions, embedding the user's original query and the structured representation of the evidence graphs into the prompt word template; Step 5.2: Use the large language model Qwen3 to generate the answer. During the generation process, record the source of evidence corresponding to each part of the answer, and label each key argument of the answer with the identifier of supporting evidence to achieve answer-evidence alignment mapping. Step 5.3: Retrieve - Reason - Generate Closed-Loop Feedback.

[0018] Furthermore, in step 5.3, Step 5.3.1: Calculate the fluency score, completeness score, and evidence coverage of the generated answer, check whether there is any content in the answer that contradicts the evidence diagram, and conduct a preliminary assessment of the generation quality; Step 5.3.2: When the answer quality is substandard or the evidence is insufficient, trigger the supplementary retrieval mechanism, analyze the weak points in the answer, generate targeted supplementary queries, return to step 2 to execute the supplementary retrieval, and integrate the newly retrieved evidence into the existing evidence subgraph; Step 5.3.3: Re-execute the evidence graph construction and reasoning of steps 3 to 4 using supplementary evidence, and regenerate the answer based on the updated evidence graph.

[0019] Furthermore, step 6 specifically includes: Step 6.1: Fact consistency verification, which is divided into direct verification using the knowledge graph and verification using the chain of evidence reasoning; Step 6.1.1: Direct verification using the knowledge graph. The generated answer is decomposed into atomic fact triples. Each triple is queried in the original knowledge graph to see if it exists. Facts that can be directly verified and those that cannot be verified are marked. The direct verification pass rate is calculated as the first-level confidence index. Step 6.1.2: Evidence chain reasoning verification. For facts that cannot be directly verified, check whether they can be derived through evidence chain reasoning. Trace the reasoning path of each argument in the answer, verify the validity of the reasoning logic, and calculate the reasoning verification pass rate as the second-level confidence index. Step 6.2: Dual-channel cross-validation, which consists of channel consistency verification and redundancy verification strategies, and finally conflict detection and resolution; Step 6.2.1: Channel consistency check. For the key facts in the answer, use the semantic retrieval channel and the structured retrieval channel to independently verify them respectively. Compare the verification results returned by the two channels and calculate the degree of consistency. Step 6.2.2: Design a redundancy verification strategy. For the core argument, there must be at least two independent evidence paths to support it. Verify whether different evidence paths lead to the same conclusion. Use evidence diversity and evidence redundancy as reliability indicators. Step 6.2.3: Identify contradictory information or conflicting reasoning paths in the evidence diagram, and resolve conflicts based on the authority, timeliness, and support of the evidence source; for conflicts that cannot be resolved, clearly state in the answer that there are different viewpoints or uncertainties. Step 6.3: Comprehensive confidence assessment and output, including multi-dimensional confidence fusion, confidence grading, and final answer output; Step 6.3.1: Integrate the confidence indices obtained in steps 6.1 and 6.2, and use a weighted fusion method to calculate the overall confidence score of the answer; Step 6.3.2: Classify the answer content according to the overall confidence score into high confidence, medium confidence, and low confidence; Step 6.3.3: Decide whether to output an answer based on the confidence assessment results. For answers with high confidence, output the complete content directly. For answers containing low-confidence content, remove or mark the uncertain parts. When the overall confidence is too low, trigger the supplementary retrieval mechanism in step 5.3 or inform the user that a reliable answer cannot be provided.

[0020] A dual-channel retrieval-guided evidence graph reasoning enhancement generation system; The system includes: an index system construction module, an adaptive dual-channel collaborative retrieval module, an evidence expansion module, an evidence graph construction module, an answer generation module, and a verification module; The index system construction module is used to receive and parse multimodal domain heterogeneous knowledge graphs, construct a pattern layer representation, perform multi-granularity embedding representation learning at the entity level, relation level and path level respectively, and establish a hybrid index system and mapping synchronization mechanism of graph database and vector database. The adaptive dual-channel collaborative retrieval module performs intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering, and the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database; The evidence expansion module performs weighted fusion and consistency / cross-validation on the two-channel retrieval results to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, the evidence graph is expanded and pruned to obtain a compact and high-quality evidence subgraph. The evidence graph construction module performs graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serializes and encodes candidate paths and calculates path confidence, and selects Top-M high-quality reasoning paths; it then transforms the reasoning paths into structured evidence chains and performs confidence propagation and aggregation. The answer generation module represents the evidence subgraph and reasoning path in a structured and sequential manner, constructs a unified prompt word framework that includes "query restatement - evidence presentation - generation instruction", guides the large language model to generate the answer, and records the evidence identifiers of key arguments to achieve answer-evidence alignment; when the generation quality is substandard, it triggers supplementary retrieval and returns to the evidence construction and reasoning process to form a closed loop; The verification module performs multiple verifications and confidence assessments on the generated results, including direct verification of the knowledge graph, verification of evidence chain reasoning, dual-channel consistency test, verification of redundant evidence and conflict detection and resolution, and obtains a comprehensive confidence level based on weighted fusion, which determines whether to output directly, label uncertainty, or trigger closed-loop error correction.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the problems of single retrieval granularity, opaque reasoning, and difficulty in verifying generated results in existing retrieval-enhanced generation technologies. It achieves precise understanding of query intent and multi-granularity knowledge acquisition through a dual-channel retrieval mechanism; it improves the interpretability of reasoning by organizing discrete knowledge into structured reasoning evidence through evidence graph construction technology; and it ensures the accuracy and reliability of generated results through multiple verification mechanisms. The overall method has good versatility and can be widely applied in various fields such as intelligent question answering, knowledge reasoning, and decision support, providing an efficient, reliable, and interpretable solution for knowledge-enhanced generation technology. Attached Figure Description

[0022] Figure 1 This is a system architecture diagram of the dual-channel retrieval-guided evidence graph reasoning enhancement generation method of the present invention; Figure 2 This is a diagram of the hybrid indexing architecture of the present invention; Figure 3 This is a schematic diagram of the multi-hop reasoning path of the present invention; Figure 4 This is a schematic diagram of the dual-channel fusion strategy of the present invention. Detailed Implementation

[0023] 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.

[0024] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0025] like Figure 1 This is a system architecture diagram of the dual-channel retrieval-guided evidence graph reasoning enhancement generation method of the present invention; the present invention provides a dual-channel retrieval-guided evidence graph reasoning enhancement generation method, specifically including the following steps: Step 1: Multimodal knowledge graph reception and multi-level indexing system construction. This involves receiving and parsing heterogeneous (potentially multimodal) knowledge graphs from various domains, constructing a schema layer representation, and performing multi-granularity embedding representation learning at the entity, relation, and path levels. A hybrid indexing system and mapping synchronization mechanism combining a graph database (Neo4j) and a vector database (Milvus) are also established. Specifically, this includes the following steps: Step 1.1: Heterogeneous Knowledge Graph Reception and Parsing. Receive complete domain knowledge graph data, including entity sets, relation sets, attribute sets, and their corresponding metadata. The knowledge graph can be formally represented as a seven-tuple:

[0026] in E = { e 1, e 2, ..., e n} represents a set of entities. n The total number of entities; R = { r 1, r 2, ..., r m} represents a set of relations. m Number of relation types; A = { a 1, a 2, ..., a k} represents a collection of attributes;T E : E → C E For entity type mapping functions, C E A hierarchical tree of entity types; T R : R → C R This is a semantic category mapping function for relations; φ : E × A → V Functions for assigning values ​​to entity attributes V For attribute value range; ψ : E × R × E → [0,1] represents the triple confidence function. The knowledge graph is structured and parsed to identify metadata such as entity type hierarchy, relation type semantic categories, and attribute constraint rules, thus constructing a schema layer representation of the knowledge graph.

[0027] Table 1. Formal Definitions and Symbol Explanations of Knowledge Graphs

[0028] Step 1.2: Multi-granularity graph embedding representation learning. A multi-level graph representation learning strategy is adopted to capture both local structural features and global semantic information. Specifically, it includes the following steps: Step 1.2.1: Entity-level embedding generation. The domain-adaptive language model BGE-Embedding is used to encode entity names, aliases, and descriptive text, generating context-aware semantic vectors. For entities... e Its text description is D ( e )= [ name ( e ); alias ( e ); desc ( e This concatenates the entity name, alias, and description text. Entity embedding is calculated as follows: he text = BGE-Encoder(D(e)) ∈ d in d = 768 represents the embedding dimension. A contrastive learning mechanism is introduced to enhance the discriminative power of entity embeddings through comparative training with synonymous and heteronymous entity pairs. The contrastive loss function is defined as: LCL = -log[exp(sim(hi , h i + ) / τ) / Σj exp(sim(h i , h j ) / τ)] in h i + For entities e i Synonymous entity embedding, τ = 0.07 is the temperature parameter. sim (·,·) is the cosine similarity function.

[0029] Entity-level embedding generation uses the domain-adaptive language model BGE-Embedding to encode entity names, aliases, and descriptive text, generating context-aware semantic vectors. A contrastive learning mechanism is introduced to enhance the discriminativeness of entity embeddings through comparative training of synonymous and heteronymous entity pairs.

[0030] Step 1.2.2: Relation-aware structured embedding, using the RotatE rotational knowledge graph embedding method to learn the representation of entities in the relation space. For triples ( h , r , t RotatE models relations as rotations in complex space: t = h r in Represents the Hadamard product. r = e (iθr) Let be the rotation vector corresponding to the relation. The scoring function is: f(h, r, t) = -||h r - t|| CompGCN is used to fuse relationship type information and aggregate the features of neighbor nodes and relationships through a message passing mechanism.

[0031] Step 1.2.3: Path-level semantic embedding to identify meta-path patterns in the knowledge graph. A meta-path is defined as an alternating sequence of entity types and relation types: P = T 1 → (R 1 ) T 2 → (R 2 ) ... → (R l ) T (l+1) The metapath is sequence-encoded using a Transformer Encoder to learn a vector representation of the multi-hop inference pattern:

[0032] Path embedding h P Retrieve the [CLS] position vector output by the Transformer. Build a path pattern library. Ω = { P 1, P 2,..., P K}, K The number of predefined path patterns provides template support for subsequent reasoning.

[0033] Table 2. Multi-granularity embedding representation parameter configuration

[0034] Step 1.3: Constructing a hybrid index system, specifically including the following steps: Figure 2 This diagram illustrates the hybrid indexing architecture, showcasing the collaborative working mechanism between the Neo4j graph database and the Milvus vector database. Step 1.3.1: For the graph database index layer, use Neo4j to build native graph storage, establish node attribute indexes and relation type indexes, and create composite indexes to support multi-condition filtering queries.

[0035] Step 1.3.2: For the vector database index layer, entity embeddings, relation embeddings, and path embeddings are stored in different collections of the Milvus vector database. A hierarchical navigation small world graph (HNSW) index is built for each collection, and a multi-level index structure is set up to balance retrieval speed and recall. The HNSW index parameters are configured as follows.

[0036] Table 3 HNSW Index Parameter Configuration

[0037] Step 1.3.3: Establish a mapping table between the graph database and the vector database, associate the structured representation and vector representation of the same entity with unique identifiers, and design an index synchronization update strategy so that when the knowledge graph is incrementally updated, the data of the two index layers are updated synchronously.

[0038] Step 2: After completing Step 1, design an adaptive dual-channel collaborative retrieval mechanism to achieve efficient and accurate retrieval. The design incorporates an adaptive dual-channel collaborative retrieval system: It performs intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering; and the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database. Specifically, it includes the following steps: Step 2.1: Understanding and decomposing the query intent, specifically including the following steps: Step 2.1.1: Query type identification. The intent of user queries is classified using a large language model based on Prompt Engineering, which includes factual queries, inference queries, multi-hop relational queries, and aggregate statistical queries. Structured information such as key entities, relational constraints, attribute filtering conditions, and time limits in the query is extracted.

[0039] Table 4. Query Type Classification Description

[0040] Step 2.1.2: Query Expansion and Rewriting. Query expansion is performed using the domain ontology to add semantically relevant candidate terms. Let the original query be... Q The expanded query set is: Q expand = Q ∪ {q' | w∈Q, w'∈Synonym(w), q'=Q[w→w']} The query was rewritten using Qwen3, generating multiple semantically equivalent but differently expressed query variants, and disambiguation was performed on fuzzy queries.

[0041] Step 2.2: Semantic retrieval channel design, which includes the following steps: Step 2.2.1: Multi-granular semantic matching. The query is fine-grained, and semantic representations at the word, phrase, and sentence levels are extracted. Multi-granular retrieval is performed in the vector database. For entity-level retrieval, the Top-K1 entity nodes most semantically similar to the query are recalled. The entity-level retrieval similarity is calculated as follows: S entity(q, e) = cos(h q , h e text ) For relational retrieval, recalling and querying Top-ranked relational patterns is crucial. K Two relation types; for path-level retrieval, the top-recall and query inference chains are similar. K Three metapath patterns.

[0042] Step 2.2.2: Post-retrieval re-ranking. The ColBERTv2 interaction model is used to refine and re-rank the candidate results. The maximum similarity score between each token in the query and the token in the candidate entity description is calculated, and the scores are accumulated to obtain the overall relevance score.

[0043] in q i For the query i A vector representation of each token. d j For document d The Middle j A vector representation of each token.

[0044] Step 2.3: Structured retrieval channel design, which includes the following steps: Step 2.3.1: Query statement generation. The natural language query is converted into a Cypher graph query language, enabling the large language model to generate structured query statements. For example, for the query "Find all people who work in City N and graduated from University M", the generated Cypher statement is: MATCH (p:Person)-[:WORK_AT]->(c:Company)-[:LOCATED_IN]->(l:Location{name:'N City'}), (p)-[:GRADUATE_FROM]->(u:University {name:'M University'}) RETURN p The generated query statements are then subjected to syntax validation and logical consistency checks.

[0045] Step 2.3.2: Constraint filtering. Perform exact matching queries with multiple constraints in the graph database, supporting complex query conditions such as attribute range filtering, relation path length limits, and subgraph pattern matching.

[0046] Step 2.3.3: Extract local subgraphs from the matching results, including nodes that satisfy the constraints and their neighboring nodes, and retain the topological structure information and edge weight information of the subgraphs.

[0047] Step 2.4: Dual-channel fusion and evidence node screening, specifically including the following steps: Step 2.4.1: Fusion Strategy Design. A weighted fusion method is adopted, dynamically adjusting the weight coefficients of the two channels according to the query type. For factual queries, the weight of the structured channel is increased; for open reasoning queries, the weight of the semantic channel is increased. A ranking learning model is used to learn the optimal fusion weights. The fusion score is calculated using the following formula: S fusion (e) = α · S semantic (e) + (1-α) · S structural (e) Step 2.4.2: Consistency verification and cross-validation. Calculate the intersection nodes of the retrieval results of the two channels as high-confidence evidence seeds. Perform cross-validation on nodes that appear only in one channel, check their relevance scores in the other channel, mark conflicting nodes, and provide uncertainty information for subsequent reasoning.

[0048] Step 2.4.3: Initial screening and ranking of evidence nodes. Candidate nodes are scored based on multiple dimensions, including semantic similarity, structural matching, and node importance. The scores are then normalized using softmax to generate the confidence distribution of the evidence nodes.

[0049] The top-N high-confidence nodes were selected as the seed node set for constructing the evidence graph. The experiment was conducted with the following settings: N = 20.

[0050] Step 3: After completing Step 2, perform weighted fusion and consistency / cross-validation on the retrieval results of the two channels to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, perform evidence graph expansion and pruning (including attention neighbor sampling, meta-path guided expansion and reinforcement learning exploration, etc.) to obtain a compact and high-quality evidence subgraph; The set of seed nodes used to construct the evidence graph needs to be expanded into the evidence graph itself. The specific process is as follows: Step 3.1: Use the evidence nodes selected in Step 2 as the seed node set. S = { s 1, s 2, ..., s n Each seed node is assigned an initial activation value, the size of which is proportional to the node's confidence level. The initial activation value is assigned based on the fusion confidence level.

[0051] Where η is the activation value scaling factor, and in the experiment, η is set to 1.0. The activation values ​​satisfy the normalization constraint:

[0052] Step 3.2: Design a multi-strategy graph expansion mechanism to expand the initial seed node, which includes the following steps: Step 3.2.1: Attention-based neighbor sampling. For each seed node, the importance weights of its neighbor nodes are calculated using a graph attention mechanism (GAT). The sampling probabilities of different types of neighbors are adjusted according to the semantic relevance of the relation type. A Top-K sampling strategy is adopted to select the K most relevant neighbors for each node for expansion. For seed node s i and its neighboring node v j ∈ N(s i The attention coefficient is calculated as follows:

[0053] where W ∈ (d'×d) To share the linear transformation matrix, W Embedding transformation matrix for relation types, r ij For connection s i and v j The relation type, a ∈ (3d') Let be the attention vector, and || denote the vector concatenation operation.

[0054] Step 3.2.2: Meta-path-guided expansion. Using the meta-path pattern library built in Step 1, the applicable inference path templates for the query are identified, and targeted expansion is performed along the predefined meta-paths to ensure that the expanded subgraph contains meaningful inference links. Multi-hop expansion is supported, and the expansion depth is adaptively adjusted according to the query complexity.

[0055] Define a meta-path matching score function to measure the fit between a query and a predefined meta-path pattern:

[0056] in For the metapath P k Embedded vector, To query the estimated inference depth, Let λ be the path length, and λ = 0.3 be the length penalty coefficient.

[0057] When expanding along the metapath, the node expansion probability is calculated as follows:

[0058] in [·] is an indicator function, T next (P) represents the expected entity type for the next step in the metapath.

[0059] Step 3.2.3: Reinforcement learning-driven exploration strategy. The graph expansion process is modeled as a Markov decision process MDP = (S, A, T, R, γ), and a deep Q-network is used to learn the optimal expansion strategy. The current evidence subgraph state is represented as:

[0060] in For the pooled representation of the current subgraph, f t = [|V t |, |E t |, d av g, coverage t [] represents the statistical feature vector of the subgraph.

[0061] The action in the action space is defined as selecting the next extended node:

[0062] Where C t This represents the current set of candidate expansion nodes. The Q network employs a two-layer MLP combined with graph convolution:

[0063] Deep Q-Network loss function:

[0064] Where D is the experience playback buffer, θ - These are the target network parameters.

[0065] Step 3.3: Refining and pruning the evidence subgraph to generate the evidence graph, which includes the following steps: Step 3.3.1: Node Importance Assessment. GraphSAINT, a graph neural network, is used to learn the representation of the expanded subgraph, calculating the centrality index of each node. This is then combined with the PageRank algorithm to evaluate the global importance of each node in the subgraph and its relevance to the query. Mixed Importance Score:

[0066] Relevance score between node and query:

[0067] Step 3.3.2: Path quality scoring. Taking into account path length, semantic coherence of relationships along the path, and relevance of the endpoint node, all paths originating from the seed node are scored. Path quality scoring function:

[0068] Where μ = 0.1 is the length attenuation coefficient. This represents the path length.

[0069] Semantic coherence is calculated using the semantic similarity of adjacent relationships:

[0070] Step 3.3.3: Identify and remove redundant nodes and edges that are irrelevant to the query. Use the Louvain community detection algorithm to identify the module structure in the subgraph, retain the core communities that are relevant to the query, set a relevance threshold, and prune nodes and edges with scores below the threshold to ensure the compactness and high quality of the evidence subgraph.

[0071] Step 3.3.4: For complex queries, generate multiple evidence subgraphs from different perspectives. Each subgraph focuses on a different reasoning dimension or interpretation angle to support the generation of diverse answers in the future.

[0072] Step 4: Perform graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serialize and encode candidate paths and calculate path confidence, and screen Top-M high-quality reasoning paths; further, transform the reasoning paths into structured evidence chains and perform confidence propagation and aggregation; After completing step 3, it is necessary to perform multi-path evidence chain mining and evidence graph construction, which includes the following steps: Step 4.1: Graph Neural Network Inference, which specifically includes the following steps: Step 4.1.1: Model the evidence subgraph using R-GCN, assigning independent embedding spaces and transformation matrices to different types of nodes and edges, aggregating neighbor information through multi-layer message passing, and updating the hidden state of nodes. For node i, the hidden state update at layer l is as follows:

[0073] in Let i be the set of neighbors of node i under relation r. = The normalization constant is This is the self-connection weight matrix. Basis decomposition is used to reduce the number of parameters.

[0074] Where B is the number of bases, For relation r to base The coefficient.

[0075] Step 4.1.2: Introduce cross-layer connections and attention mechanisms into the multi-layer GNN, enabling each layer's node representation to receive information from other layers, capturing structural features at different scales. Attention weights are used to dynamically fuse representations from different layers to generate the final node embedding. With the introduction of cross-layer connections, the final representation of layer L is:

[0076] The cross-layer attention weights are calculated through query-guided calculation.

[0077] Step 4.1.3: Inject the query vector as additional input into each layer of the GNN, using a gating mechanism to adjust the impact of the query information on the node representation, and learn subgraph representations highly relevant to the query. Inject the query vector into each layer of the GNN:

[0078]

[0079] Step 4.2: Multi-hop reasoning path discovery. Identify multiple reasoning paths from the query entity to the target entity in the evidence subgraph, such as... Figure 3 This is a diagram illustrating multi-hop inference paths, showing multiple inference paths from the query entity to the answer entity and their confidence levels; specifically, it includes the following steps: Step 4.2.1: Enumerate all feasible paths from the query-related entity to the potential answer entity in the evidence subgraph, and use depth-first search combined with heuristic pruning strategy to enumerate candidate paths.

[0080] Step 4.2.2: Represent each path as an alternating sequence of entity-relation relationships and encode the path sequence, introducing positional encoding and relation type embedding to enhance the semantic expressiveness of the path and achieve path semantic modeling. Introducing relation type positional encoding:

[0081] Step 4.2.3: Calculate the confidence score for each path. Taking into account path length penalty, semantic coherence of relations, and entity relevance, use a ranking learning model to rank the paths, retaining the Top-M high-quality inference paths. Path confidence score:

[0082] Training a sorting model using the ListMLE loss function:

[0083] Step 4.3: Generate the inference evidence chain. Transform the high-confidence inference path into a structured evidence chain, and label the relationship type and confidence score of each edge. This includes the following steps: Step 4.3.1: Transform the filtered reasoning paths into a structured chain of evidence, label each edge with the relationship type, relationship strength, and number of supporting evidence, construct a hierarchical structure of the chain of evidence, and distinguish between the main reasoning path and the auxiliary supporting path. Step 4.3.2: Use a probabilistic graphical model to model the confidence of the evidence chain. Starting from the seed node, propagate the confidence along the path, considering the reliability of the edges and the prior confidence of the nodes, and aggregate the confidence of multiple evidence chains.

[0084] Step 5: Represent the evidence subgraph and reasoning path in a structured and sequential manner, construct a unified prompt word framework containing "query restatement - evidence presentation - generation instruction", guide the large language model to generate the answer, and record the evidence identifiers of key arguments to achieve answer-evidence alignment; when the generation quality is substandard, trigger supplementary retrieval and return to the evidence construction and reasoning process to form a closed loop; After completing step 4, you need to generate an answer guided by evidence diagrams, which includes the following steps: Step 5.1: Structured Representation of Evidence Graphs. The evidence subgraphs and reasoning paths constructed in Steps 3 and 4 are transformed into structured text representations, enabling the serialization of evidence subgraphs. A unified prompt word framework is designed, comprising three core parts: query restatement, evidence presentation, and instruction generation. The user's original query and the structured representation of the evidence graph are embedded into the prompt word template.

[0085] Step 5.2: Use the large language model Qwen3 to generate the answer. During the generation process, record the source of evidence corresponding to each part of the answer, and label each key argument of the answer with the identifier of supporting evidence to achieve answer-evidence alignment mapping.

[0086] Step 5.3: Retrieval-Inference-Generation of Closed-Loop Feedback, specifically including the following steps: Step 5.3.1: Calculate the fluency score, completeness score, and evidence coverage of the generated answer, check whether there is any content in the answer that contradicts the evidence diagram, and conduct a preliminary assessment of the quality of the generated answer.

[0087] Fluency score:

[0088] Integrity score: Extracting a set of key points from a query Extracting a set of key points from the answer ,definition:

[0089] Evidence coverage: Extracting atomic claims from the answers If you advocate Alignment evidence exists Then it is denoted as a cover:

[0090] Conflict rate: If assertion A conflict between the evidence and its alignment (which can be verified using NLI / rule validation) is denoted as a contradiction:

[0091] Step 5.3.2: When the answer quality is substandard or the evidence is insufficient, trigger the supplementary retrieval mechanism, analyze the weak points in the answer, generate targeted supplementary queries, return to step 2 to execute the supplementary retrieval, and integrate the newly retrieved evidence into the existing evidence subgraph.

[0092] Step 5.3.3: Re-execute the evidence graph construction and reasoning of Step 3-4 using supplementary evidence, and regenerate the answer based on the updated evidence graph.

[0093] Step 6: Perform multiple verifications and confidence assessments on the generated results: including direct verification from the knowledge graph, verification through evidence chain reasoning, dual-channel consistency testing, verification of redundant evidence, and conflict detection and resolution. Based on weighted fusion, a comprehensive confidence level is obtained to determine whether to output directly, label uncertainties, or trigger closed-loop error correction.

[0094] like Figure 4 This is a schematic diagram of a dual-channel fusion strategy, illustrating the collaborative fusion process of the semantic and structured channels. After completing step 5, a multi-verification mechanism and confidence assessment need to be designed, which specifically includes the following steps: Step 6.1: Fact consistency verification, which is divided into direct verification using the knowledge graph and verification using the chain of evidence reasoning. Specifically, it includes the following steps: Step 6.1.1: Direct verification using the knowledge graph. The generated answer is decomposed into atomized fact triples. Each triple is queried in the original knowledge graph to see if it exists. Facts that can be directly verified and facts that cannot be verified are marked. The direct verification pass rate is calculated as the first-level confidence index.

[0095] KG direct verification pass rate (if the triple can be matched in the original knowledge graph, it is counted as pass):

[0096] Step 6.1.2: Evidence Chain Reasoning Verification. For facts that cannot be directly verified, check whether they can be derived through evidence chain reasoning. Trace the reasoning path of each argument in the answer, verify the validity of the reasoning logic, and calculate the reasoning verification pass rate as the second-level confidence index. The evidence chain reasoning verification pass rate is for the subset that cannot be directly verified. If there is an evidentiary path / rule that leads to this fact, then the reasoning is considered successful.

[0097] Step 6.2: Dual-channel cross-validation, which consists of channel consistency verification and redundancy verification strategies, and finally conflict detection and resolution. Specifically, it includes the following steps: Step 6.2.1: Channel Consistency Check. For the key facts in the answer, independently verify them using both the semantic retrieval channel and the structured retrieval channel. Compare the verification results returned by the two channels and calculate the degree of consistency. The channel consistency degree is related to the set of key facts. The semantic channel and the structured channel respectively provide the verification results. :

[0098] Step 6.2.2: Design a redundancy verification strategy. For the core argument, at least two independent evidence paths are required to support it, and it is verified whether different evidence paths lead to the same conclusion. Evidence diversity and evidence redundancy are used as reliability indicators. The redundancy verification indicators apply to the set of core arguments. Let the number of independent evidence paths be . :

[0099] Step 6.2.3: Identify contradictory information or conflicting reasoning paths in the evidence diagram, and resolve the conflicts based on the authority, timeliness, and support of the evidence sources. For conflicts that cannot be resolved, clearly state in the answer that there are differing viewpoints or uncertainties.

[0100] Step 6.3: Comprehensive confidence assessment and output, including multi-dimensional confidence fusion, confidence grading, and final answer output, specifically including the following steps: Step 6.3.1: Integrate the confidence indices obtained in Steps 6.1 and 6.2, and calculate the overall confidence score of the answer using a weighted fusion method. Overall Confidence Score:

[0101] in For conflict rate.

[0102] Step 6.3.2: Classify the answer content according to the overall confidence score into high confidence, medium confidence, and low confidence.

[0103]

[0104] Step 6.3.3: Decide whether to output an answer based on the confidence assessment results. For answers with high confidence, output the complete content directly. For answers containing low-confidence content, remove or mark the uncertain parts. When the overall confidence is too low, trigger the supplementary retrieval mechanism in step 5.3 or inform the user that a reliable answer cannot be provided.

[0105] A dual-channel retrieval-guided evidence graph reasoning enhancement generation system; The system includes: an index system construction module, an adaptive dual-channel collaborative retrieval module, an evidence expansion module, an evidence graph construction module, an answer generation module, and a verification module; The index system construction module is used to receive and parse multimodal domain heterogeneous knowledge graphs, construct a pattern layer representation, perform multi-granularity embedding representation learning at the entity level, relation level and path level respectively, and establish a hybrid index system and mapping synchronization mechanism of graph database and vector database. The adaptive dual-channel collaborative retrieval module performs intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering, and the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database; The evidence expansion module performs weighted fusion and consistency / cross-validation on the two-channel retrieval results to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, the evidence graph is expanded and pruned to obtain a compact and high-quality evidence subgraph. The evidence graph construction module performs graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serializes and encodes candidate paths and calculates path confidence, and selects Top-M high-quality reasoning paths; it then transforms the reasoning paths into structured evidence chains and performs confidence propagation and aggregation. The answer generation module represents the evidence subgraph and reasoning path in a structured and sequential manner, constructs a unified prompt word framework that includes "query restatement - evidence presentation - generation instruction", guides the large language model to generate the answer, and records the evidence identifiers of key arguments to achieve answer-evidence alignment; when the generation quality is substandard, it triggers supplementary retrieval and returns to the evidence construction and reasoning process to form a closed loop; The verification module performs multiple verifications and confidence assessments on the generated results, including direct verification of the knowledge graph, verification of evidence chain reasoning, dual-channel consistency test, verification of redundant evidence and conflict detection and resolution, and obtains a comprehensive confidence level based on weighted fusion, which determines whether to output directly, label uncertainty, or trigger closed-loop error correction.

[0106] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0107] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0108] The above provides a detailed description of the dual-channel retrieval-guided evidence graph reasoning enhancement generation method proposed in this invention, and elucidates the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A dual-channel retrieval-guided method for enhancing evidence graph reasoning generation, characterized in that: The method specifically includes the following steps: Step 1: Receive and parse multimodal domain heterogeneous knowledge graphs containing text, images, videos, and audio, and construct a pattern layer representation; perform multi-granularity embedding representation learning at the entity level, relation level, and path level respectively, and establish a hybrid indexing system and mapping synchronization mechanism for graph databases and vector databases; Step 2, design adaptive dual-channel collaborative retrieval: perform intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering, while the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database; Step 3: Weighted fusion and verification of the two-channel retrieval results are performed to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, evidence graph expansion and pruning are performed to obtain a compact and high-quality evidence subgraph. Step 4: Perform graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serialize and encode candidate paths and calculate path confidence, and screen Top-M high-quality reasoning paths; transform the reasoning paths into structured evidence chains and perform confidence propagation and aggregation. Step 5: The evidence subgraph and reasoning path are represented in a structured and sequential manner. A unified prompt word framework containing "query restatement - evidence presentation - generation instruction" is constructed to guide the large language model to generate the answer and record the evidence identifiers of key arguments to achieve answer-evidence alignment. When the generation quality is not up to standard, a supplementary retrieval is triggered and the process of evidence construction and reasoning is returned to form a closed loop. Step 6: Perform multiple verifications and confidence assessments on the generated results, including direct verification using the knowledge graph, verification through evidence chain reasoning, dual-channel consistency testing, verification of redundant evidence, and conflict detection and resolution. Based on weighted fusion, obtain a comprehensive confidence level and decide whether to output directly, label uncertainties, or trigger closed-loop error correction.

2. The method according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Heterogeneous knowledge graph reception and parsing. Receive complete domain knowledge graph data, including entity sets, relation sets, attribute sets, and their corresponding metadata information; The knowledge graph is structured and parsed to identify meta-information including hierarchical relationships of entity types, semantic categories of relationship types, and attribute constraint rules, and to construct a schema layer representation of the knowledge graph. Step 1.2: Multi-granularity graph embedding representation learning, adopting a multi-level graph representation learning strategy to capture both local structural features and global semantic information; Step 1.3: Construct a hybrid indexing system; For the graph database index layer, Neo4j is used to build native graph storage, node attribute indexes and relation type indexes are created, and composite indexes are created to support multi-condition filtering queries. For the vector database index layer, entity embedding, relation embedding, and path embedding are stored in different sets of the Milvus vector database. A hierarchical navigation small world graph (HNSW) index is built for each set, and a multi-level index structure is set to balance retrieval speed and recall. Establish a mapping table between the graph database and the vector database, and associate the structured representation and vector representation of the same entity through a unique identifier. At the same time, design an index synchronization update strategy to update the data of the two index layers synchronously when the knowledge graph is incrementally updated.

3. The method according to claim 2, characterized in that: Step 1.2 specifically includes: Step 1.2.1: Entity-level embedding generation. The domain-adaptive language model BGE-Embedding is used to encode entity names, aliases, and descriptive text to generate context-aware semantic vectors. A contrastive learning mechanism is introduced to enhance the discriminativeness of entity embeddings through comparative training of synonymous entity pairs and heteronymous entity pairs. Step 1.2.2: Relationship-aware structured embedding. The Rotat rotating knowledge graph embedding method is used to learn the representation of entities in the relation space. CompGCN is used to fuse relation type information. The features of neighbor nodes and relations are aggregated through message passing mechanism. Weight matrices are assigned to different relation types to capture the semantic differences of heterogeneous relations. Step 1.2.3: Path-level semantic embedding, identify meta-path patterns in the knowledge graph, use TransformerEncoder to sequence encode the meta-paths, learn the vector representation of multi-hop reasoning patterns, and build a path pattern library.

4. The method according to claim 3, characterized in that: Step 2 specifically includes: Step 2.1: Understanding and decomposing the query intent; Step 2.1.1: Query type identification. The user query intent is classified using a large language model based on Prompt Engineering, which is divided into factual queries, inference queries, multi-hop relational queries, and aggregate statistical queries. The structured information in the query is extracted, including key entities, relational constraints, attribute filtering conditions, and time limits. Step 2.1.2: Query expansion and rewriting. The query is expanded using the domain ontology to add semantically related candidate words. Qwen3 is used to rewrite the query to generate multiple semantically equivalent but different query variants. Disambiguation is performed on fuzzy queries. Step 2.2: Semantic retrieval channel design; Step 2.2.1: Multi-granular semantic matching. The query is encoded in a fine-grained manner, and semantic representations at the word, phrase, and sentence levels are extracted respectively. Multi-granular retrieval is performed in the vector database. For entity-level retrieval, the Top-K1 entity nodes that are most similar to the query semantics are recalled; for relation-level retrieval, the Top-K2 relation types that match the query relation pattern are recalled; for path-level retrieval, the Top-K3 meta-path patterns that are similar to the query inference link are recalled. Step 2.2.2: Re-ranking after retrieval. The ColBERTv2 interaction model is used to refine and re-rank the candidate results. The maximum similarity score between each token in the query and the token in the candidate entity description is calculated, and the scores are accumulated to obtain the overall relevance score. Step 2.3: Design of structured retrieval channels; Step 2.3.1: Query statement generation. The natural language query is converted into Cypher graph query language, enabling the large language model to generate structured query statements, and the generated query statements are subjected to syntax validation and logical consistency checks. Step 2.3.2: Constraint filtering, performing exact match queries with multiple constraints in the graph database; Step 2.3.3: Extract local subgraphs from the matching results, including nodes that satisfy the constraints and their neighboring nodes, and retain the topological structure information and edge weight information of the subgraph; Step 2.4: Dual-channel fusion and evidence node screening; Step 2.4.1: Design the fusion strategy. A weighted fusion method is adopted to dynamically adjust the weight coefficients of the two channels according to the query type. For factual queries, the weight of the structured channel is increased; for open reasoning queries, the weight of the semantic channel is increased. The optimal fusion weight is learned using a ranking learning model. Step 2.4.2: Consistency verification and cross-validation. Calculate the intersection nodes of the retrieval results of the two channels as high-confidence evidence seeds. Perform cross-validation on nodes that appear only in one channel, check their relevance scores in the other channel, mark conflicting nodes, and provide uncertainty information for subsequent reasoning. Step 2.4.3: Initial screening and sorting of evidence nodes. Based on semantic similarity, structural matching degree and node importance, candidate nodes are scored. The scores are processed using softmax normalization to generate the confidence distribution of evidence nodes. The top-N high-confidence nodes are selected as the seed node set for the construction of the evidence graph.

5. The method according to claim 4, characterized in that: Step 3 specifically includes: Step 3.1: Using the evidence nodes selected in Step 2 as the seed node set, assign an initial activation value to each seed node. The activation value is proportional to the confidence level of the node. Step 3.2: Design a multi-strategy graph expansion mechanism to expand the initial seed node; Step 3.2.1: Attention-based neighbor sampling. For each seed node, the importance weight of its neighbor nodes is calculated using a graph attention mechanism. The sampling probability of different types of neighbors is adjusted according to the semantic relevance of the relation type. A Top-K sampling strategy is adopted to select the K most relevant neighbors for each node for expansion. Step 3.2.2: Meta-path-guided expansion. Using the meta-path pattern library built in Step 1, identify the inference path templates applicable to the query and expand in a targeted manner along the predefined meta-paths. Step 3.2.3: Reinforcement learning-driven exploration strategy, modeling the graph expansion process as a Markov decision process, and using a deep Q-network to learn the optimal expansion strategy; Step 3.3: Refine and prune the evidence subgraph to generate the evidence graph; Step 3.3.1: Node importance assessment. GraphSAINT, a graph neural network, is used to learn the representation of the expanded subgraph. The centrality index of each node is calculated and the PageRank algorithm is used to evaluate the global importance of the node in the subgraph and its relevance to the query. Step 3.3.2: Path quality scoring, taking into account path length, semantic coherence of relationships on the path, and relevance of the endpoint node, to score all paths starting from the seed node; Step 3.3.3: Identify and remove redundant nodes and edges that are irrelevant to the query. Use the Louvain community detection algorithm to identify the module structure in the subgraph, retain the core communities that are relevant to the query, set a relevance threshold, and prune nodes and edges with scores below the threshold. Step 3.3.4: For complex queries, generate multiple evidence subgraphs from different perspectives.

6. The method according to claim 5, characterized in that: Step 4 specifically includes: Step 4.1: Graph Neural Network Inference; Step 4.1.1: Model the evidence subgraph using R-GCN, assign independent embedding spaces and transformation matrices to different types of nodes and edges, aggregate neighbor information through multi-layer message passing, and update the hidden state of nodes; Step 4.1.2: Introduce cross-layer connections and attention mechanisms in the multi-layer GNN so that the node representation of each layer can receive information from other layers, capture structural features at different scales, and use attention weights to dynamically fuse the representations of different layers to generate the final node embedding. Step 4.1.3: Inject the query vector as additional input into each layer of the GNN, use a gating mechanism to adjust the influence of the query information on the node representation, and learn the subgraph representation that is highly related to the query; Step 4.2: Multi-hop reasoning path discovery, identifying multiple reasoning paths from the query entity to the target entity in the evidence subgraph; Step 4.2.1: Enumerate all feasible paths from the query-related entity to the potential answer entity in the evidence subgraph, and use depth-first search combined with heuristic pruning strategy to enumerate candidate paths; Step 4.2.2: Represent each path as an alternating sequence of entity-relationships and encode the path sequence, introducing position encoding and relation type embedding to enhance the semantic expressiveness of the path and realize path semantic modeling; Step 4.2.3: Calculate the confidence score for each path, and comprehensively consider path length penalty, semantic coherence of relations, and entity relevance. Use a ranking learning model to rank the paths and retain the Top-M high-quality inference paths. Step 4.3: Generate the inference evidence chain, transform the high-confidence inference path into a structured evidence chain, and label the relationship type and confidence score of each edge; Step 4.3.1: Transform the filtered reasoning paths into a structured chain of evidence, label each edge with the relationship type, relationship strength, and number of supporting evidence, construct the hierarchical structure of the chain of evidence, and distinguish between the main reasoning path and the auxiliary supporting path; Step 4.3.2: Use a probabilistic graphical model to model the confidence of the evidence chain. Starting from the seed node, propagate the confidence along the path, considering the reliability of the edges and the prior confidence of the nodes, and aggregate the confidence of multiple evidence chains.

7. The method according to claim 6, characterized in that: Step 5 specifically includes: Step 5.1: Structured Representation of Evidence Graphs; Transform the evidence subgraphs and reasoning paths constructed in Steps 3 and 4 into structured text representations to achieve evidence subgraph serialization; Design a unified prompt word framework, which includes three core parts: query restatement, evidence presentation, and generation instructions, embedding the user's original query and the structured representation of the evidence graphs into the prompt word template; Step 5.2: Use the large language model Qwen3 to generate the answer. During the generation process, record the source of evidence corresponding to each part of the answer, and label each key argument of the answer with the identifier of supporting evidence to achieve answer-evidence alignment mapping. Step 5.3: Retrieval - Reasoning - Generation of Closed-Loop Feedback.

8. The method according to claim 7, characterized in that: In step 5.3, Step 5.3.1: Calculate the fluency score, completeness score, and evidence coverage of the generated answer, check whether there is any content in the answer that contradicts the evidence diagram, and conduct a preliminary assessment of the generation quality; Step 5.3.2: When the answer quality is substandard or the evidence is insufficient, trigger the supplementary retrieval mechanism, analyze the weak points in the answer, generate targeted supplementary queries, return to step 2 to execute the supplementary retrieval, and integrate the newly retrieved evidence into the existing evidence subgraph; Step 5.3.3: Re-execute the evidence graph construction and reasoning of steps 3 to 4 using supplementary evidence, and regenerate the answer based on the updated evidence graph.

9. The method according to claim 8, characterized in that: Step 6 specifically includes: Step 6.1: Fact consistency verification, which is divided into direct verification using the knowledge graph and verification using the chain of evidence reasoning; Step 6.1.1: Direct verification using the knowledge graph. The generated answer is decomposed into atomic fact triples. Each triple is queried in the original knowledge graph to see if it exists. Facts that can be directly verified and those that cannot be verified are marked. The direct verification pass rate is calculated as the first-level confidence index. Step 6.1.2: Evidence chain reasoning verification. For facts that cannot be directly verified, check whether they can be derived through evidence chain reasoning. Trace the reasoning path of each argument in the answer, verify the validity of the reasoning logic, and calculate the reasoning verification pass rate as the second-level confidence index. Step 6.2: Dual-channel cross-validation, which consists of channel consistency verification and redundancy verification strategies, and finally conflict detection and resolution; Step 6.2.1: Channel consistency check. For the key facts in the answer, use the semantic retrieval channel and the structured retrieval channel to independently verify them respectively. Compare the verification results returned by the two channels and calculate the degree of consistency. Step 6.2.2: Design a redundancy verification strategy. For the core argument, there must be at least two independent evidence paths to support it. Verify whether different evidence paths lead to the same conclusion. Use evidence diversity and evidence redundancy as reliability indicators. Step 6.2.3: Identify contradictory information or conflicting reasoning paths in the evidence diagram, and resolve conflicts based on the authority, timeliness, and support of the evidence source; for conflicts that cannot be resolved, clearly state in the answer that there are different viewpoints or uncertainties. Step 6.3: Comprehensive confidence assessment and output, including multi-dimensional confidence fusion, confidence grading, and final answer output; Step 6.3.1: Integrate the confidence indices obtained in steps 6.1 and 6.2, and use a weighted fusion method to calculate the overall confidence score of the answer; Step 6.3.2: Classify the answer content according to the overall confidence score into high confidence, medium confidence, and low confidence; Step 6.3.3: Decide whether to output an answer based on the confidence assessment results. For answers with high confidence, output the complete content directly. For answers containing low-confidence content, remove or mark the uncertain parts. When the overall confidence is too low, trigger the supplementary retrieval mechanism in step 5.3 or inform the user that a reliable answer cannot be provided.

10. A dual-channel retrieval-guided evidence graph reasoning enhancement generation system, characterized in that: The system executes the dual-channel retrieval-guided evidence graph reasoning enhancement generation method as described in any one of claims 1 to 9; The system includes: an index system construction module, an adaptive dual-channel collaborative retrieval module, an evidence expansion module, an evidence graph construction module, an answer generation module, and a verification module; The index system construction module is used to receive and parse multimodal domain heterogeneous knowledge graphs, construct a pattern layer representation, perform multi-granularity embedding representation learning at the entity level, relation level and path level respectively, and establish a hybrid index system and mapping synchronization mechanism of graph database and vector database. The adaptive dual-channel collaborative retrieval module performs intent understanding, decomposition, and expansion / rewriting of user queries; the semantic retrieval channel performs multi-granularity vector recall and reordering, and the structured retrieval channel converts natural language into graph query statements and performs constraint filtering and subgraph extraction in the graph database; The evidence expansion module performs weighted fusion and consistency / cross-validation on the two-channel retrieval results to obtain a set of high-confidence evidence seed nodes and form conflict markers; on this basis, the evidence graph is expanded and pruned to obtain a compact and high-quality evidence subgraph. The evidence graph construction module performs graph neural network reasoning and multi-hop reasoning path discovery on the evidence subgraph, serializes and encodes candidate paths and calculates path confidence, and selects Top-M high-quality reasoning paths; it then transforms the reasoning paths into structured evidence chains and performs confidence propagation and aggregation. The answer generation module represents the evidence subgraph and reasoning path in a structured and sequential manner, constructs a unified prompt word framework that includes "query restatement - evidence presentation - generation instruction", guides the large language model to generate the answer, and records the evidence identifiers of key arguments to achieve answer-evidence alignment; when the generation quality is substandard, it triggers supplementary retrieval and returns to the evidence construction and reasoning process to form a closed loop; The verification module performs multiple verifications and confidence assessments on the generated results, including direct verification of the knowledge graph, verification of evidence chain reasoning, dual-channel consistency test, verification of redundant evidence and conflict detection and resolution, and obtains a comprehensive confidence level based on weighted fusion, which determines whether to output directly, label uncertainty, or trigger closed-loop error correction.

Images