Search enhancement generation method based on multi-level knowledge fusion

Abstract

This invention relates to a retrieval enhancement generation method based on multi-level knowledge fusion, belonging to the field of natural language processing technology. It aims to solve two core problems in existing retrieval enhancement generation technologies: fragmented knowledge levels and interference from noise information. The invention includes the following steps: Graph index construction: Segmenting documents into fragments and extracting entities and relations to generate a knowledge graph with key-value pairs, followed by deduplication and optimization; Multi-level knowledge retrieval: Acquiring local and global knowledge separately, and constructing cross-level semantic associations by connecting knowledge modules through an intermediate layer; Dual-dimensional filtering: Filtering core knowledge based on a weighted score of semantic relevance and knowledge importance, combined with a maximum boundary relevance strategy to remove redundant noise; Response generation: Inputting the filtered knowledge into a large-scale language model to generate accurate and coherent responses. This invention significantly improves the generation quality of knowledge-intensive tasks through cross-level knowledge bridging and precise noise filtering, possessing significant theoretical and practical application value.

Description

Technical Field

[0001] This invention relates to a retrieval enhancement generation method based on multi-level knowledge fusion, which is applicable to knowledge-intensive tasks such as Query Focused Summary (QFS) and Multi-hop Question Answering (MHQA) and belongs to the field of natural language processing technology. Background Technology

[0002] Retrieval Augmentation (RAG) technology effectively alleviates the model "illusion" problem by providing external knowledge support to large language models (LLMs), and has been widely used in specific domains and knowledge-intensive tasks. Existing RAG methods are mainly divided into two categories: one is to segment text into flat text blocks for retrieval, and the other is to organize entities and relations based on graph structures to handle global-level problems that require the integration of multiple parts of information.

[0003] However, existing RAG methods still suffer from two major drawbacks: First, fragmented knowledge hierarchy: local retrieval knowledge (entity information directly related to the query) and global retrieval knowledge (macro-topic or domain background information) are independent, lacking semantic bridging, leading to incoherent model reasoning and factual contradictions in the answers. Second, noise interference: retrieval results contain a large amount of irrelevant or redundant data, increasing model token consumption and interfering with semantic understanding, thus reducing generation quality. The positive correlation between retrieval recall and data volume indicates the presence of valuable information in the retrieval results, but distinguishing between effective knowledge and noise remains a key issue. While existing graph-based RAG methods attempt to integrate multi-source knowledge, they fail to effectively bridge the semantic gap across knowledge layers and lack precise noise filtering mechanisms, limiting their application in complex, knowledge-intensive tasks. Therefore, a retrieval enhancement generation method capable of achieving cross-layer knowledge collaboration and high-quality knowledge filtering is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a retrieval enhancement generation method based on multi-level knowledge fusion. By connecting knowledge construction and dual-dimensional filtering mechanism through an intermediate layer, it systematically solves the problems of knowledge hierarchy fragmentation and noise interference, and improves the accuracy, coherence and comprehensiveness of the generated response.

[0005] The technical solution of this invention is: a retrieval enhancement generation method based on multi-level knowledge fusion, comprising:

[0006] Step 1, Graph Index Construction: The original text document is segmented into fragments, entities and relationships between entities are extracted from the fragments, a knowledge graph with key-value pairs is generated, and the graph operation overhead is reduced by deduplication optimization;

[0007] Step 2, Multi-level Knowledge Retrieval: Extract keywords from the input query, retrieve local and global knowledge respectively, and then construct semantic associations between local and global knowledge by connecting knowledge modules through an intermediate layer;

[0008] Step 3, Two-dimensional screening: Core knowledge is screened by scoring based on relevance, and redundant information is removed by using a diversity strategy to obtain the screened knowledge set;

[0009] Step 4, Answer Generation: Input the filtered knowledge set as external knowledge into a large language model to generate the target response.

[0010] Further, step 1 includes:

[0011] Step 1.1, Document Segmentation: Divide the original text document into multiple segments;

[0012] Step 1.2, Entity and Relationship Extraction: Utilize a large language model to identify entities and relationships between entities in each segment;

[0013] Step 1.3, Key-value pair generation: Through the parsing function driven by the large language model, key-value pairs are generated for each entity node, with the entity name as the index key and the related text paragraph as the value. For each relation edge, multiple index keys containing the global topic and corresponding text values ​​are generated.

[0014] Step 1.4, Deduplication Optimization: Merge identical entities and relationships in different segments using a deduplication function to minimize the size of the knowledge graph and ultimately generate the knowledge graph.

[0015] Furthermore, the specific steps of multi-level knowledge retrieval in step 2 include:

[0016] Step 2.1, Local Knowledge Retrieval: Using LLM, extract the entity-level keyword set Kq from query Q. Through semantic embedding model, perform cosine similarity matching between Kq and nodes in knowledge graph G, retrieve a predefined number N of relevant nodes, and obtain the local knowledge set Vq.

[0017] Step 2.2, Global Knowledge Retrieval: Extract global keywords for query Q, match relationships between global keys in the knowledge graph, collect nodes and edges in the multi-hop subgraph of this relationship, and obtain the global knowledge set;

[0018] Step 2.3, Intermediate Layer Connection Knowledge Retrieval: This includes two parts: graph diffusion mechanism and path discovery.

[0019] Part 1: Graph Diffusion Mechanism: Based on a personalized PageRank algorithm, the diffusion vector P is initialized. (0) Through the iterative formula:

[0020] Update vector, where Let A be the restart probability, and A be the normalized adjacency matrix. The iteration... Subsequently, Top-k entities were selected as locally diffused entities, combined with the entity distance attenuation factor. Calculate the diffusion score;

[0021] The second part involves path discovery: constructing two types of reasoning paths R = R1 + R2, where R1 is the set of shortest paths between local knowledge entities and the head and tail entities of global knowledge relationships, and R2 is the set of paths that satisfy cross-layer relevance. Cross-layer semantic matching path, As the threshold, the formula for calculating cross-layer correlation is:

[0022]

[0023] s and t are global relations r G The head and tail entities, w(r) G ) represents the relation weight, and is selected as follows: ,turn up arrive The shortest path to the endpoint.

[0024] Furthermore, the specific steps of the two-dimensional screening in step 3 include:

[0025] Step 3.1, Relevance Dimension Filtering: A comprehensive score Rel2(k,Q) is calculated by weighted fusion of semantic relevance score and knowledge importance score. Sim(.)+ Importance(.) filters out knowledge units that are below a threshold θ. , All are weights; the semantic relevance scoring formula is:

[0026]

[0027] in, Let Q be the knowledge unit to be evaluated, and K be the query. low K is a local keyword set. high The set of global keywords, λ1+λ2=1, is dynamically set according to task characteristics; knowledge importance scoring is designed separately for entities, relationships, and document fragments.

[0028] Entity importance:

[0029]

[0030] Among them, parameters Used to balance semantic and structural features Let be the degree of entity e in the knowledge graph. For entities in a knowledge graph, yes One of the entities, and These are the embedding vectors for entity e and query Q, respectively;

[0031] Importance of the relationship: , For relation embedding vectors, For relationship Weights, parameters It is also used to balance semantic and structural features;

[0032] Importance of document fragments: ;in This is a document fragment. For document fragments The embedding vector;

[0033] Step 3.2, Diversity Dimension Filtering: Based on the maximum boundary relevance strategy, iteratively select the optimal knowledge unit.

[0034]

[0035] in, This is the set of candidate knowledge after initial screening based on relevance. This is a selected set of high-value knowledge, initially empty. The optimal knowledge unit selected for the current round, sim(k, () represents candidate knowledge k and selected knowledge The cosine similarity is used, with parameter γ∈(0,1) as a balance factor. The larger γ is, the more it emphasizes relevance, and the smaller γ is, the more it emphasizes diversity, until a preset number M knowledge units are selected to form the final knowledge set.

[0036] Furthermore, the specific steps for generating the answer in step 4 include:

[0037] Step 4.1: Using the filtered knowledge set, which includes entity name, relationship type, detailed description and key excerpts from the original text, concatenate them into the input text of "query command + knowledge support + generation requirements";

[0038] Step 4.2: Input text is passed to the target large model. The model analyzes the query requirements and knowledge associations, integrates multi-source information such as entities, relationships, and original text extracts, and establishes a precise mapping between the query and knowledge. Through logical reasoning, the language is organized according to "responding to the core question → supplementing key details" to ensure that the answer is consistent with the query. Figure 1 It provides comprehensive and accurate information, directly generating and outputting the final answer that is adapted to the task scenario.

[0039] The present invention also provides a retrieval enhancement generation system based on multi-level knowledge fusion, the system comprising: a module for executing the retrieval enhancement generation method based on multi-level knowledge fusion.

[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the retrieval enhancement generation method based on multi-level knowledge fusion.

[0041] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the retrieval enhancement generation method based on multi-level knowledge fusion.

[0042] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the retrieval enhancement generation method based on multi-level knowledge fusion.

[0043] The beneficial effects of this invention are:

[0044] 1. Solved the problem of knowledge hierarchy fragmentation: Through the graph diffusion mechanism and path discovery strategy of the intermediate layer knowledge (IKL) module, a semantic bridge between local and global knowledge is constructed, which fills the cross-layer semantic gap and ensures that the model can coordinate knowledge at different levels for coherent reasoning, avoiding the problems of factual contradictions and incoherent reasoning.

[0045] 2. Efficient filtering of noisy information: The dual-dimensional screening framework combines semantic relevance and knowledge importance scoring to accurately filter core knowledge, and then removes redundancy through the MMR strategy, which not only ensures the effectiveness of knowledge, but also controls token consumption, thereby improving the efficiency and quality of model generation.

[0046] 3. Wide applicability and excellent performance: It shows significant advantages in both query focused summarization (QFS) and multi-hop question answering (MHQA) tasks. Compared with existing mainstream baseline methods, the win rate of QFS task is improved by 5%-15%, and the exact matching rate (EM) and F1 score of MHQA task are significantly higher than existing state-of-the-art methods, which has important practical application value. Attached Figure Description

[0047] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0048] Example 1: As Figure 1 As shown, the retrieval enhancement generation method based on multi-level knowledge fusion includes:

[0049] Step 1, Graph Index Construction: The original text document is segmented into fragments, entities and relationships between entities are extracted from the fragments, a knowledge graph with key-value pairs is generated, and the graph operation overhead is reduced by deduplication optimization;

[0050] Further, step 1 includes:

[0051] Step 1.1, Document Segmentation: Segment the original text document D into multiple fragments Ds. i This facilitates quick identification of relevant information;

[0052] Step 1.2, Entity and Relation Extraction: Utilize Large Language Models (LLMs) to identify each segment D i Let V be the set of entities (nodes) and the set of relationships (edges) between them;

[0053] Step 1.3, Key-value pair generation: Using the LLM-driven parsing function P(.), generate key-value pairs for each entity node with the entity name as the index key and the related text paragraph as the value, and generate multiple index keys containing the global topic and corresponding text values ​​for each relation edge;

[0054] Step 1.4, Deduplication Optimization: Merge identical entities and relations in different segments using the deduplication function D(.) to minimize the size of the knowledge graph and improve processing efficiency; finally, generate the knowledge graph G=(V,E).

[0055] Step 2, Multi-level Knowledge Retrieval: Extract keywords from the input query, retrieve local and global knowledge respectively, and then construct semantic associations between local and global knowledge by connecting knowledge modules through an intermediate layer;

[0056] Furthermore, the specific steps of multi-level knowledge retrieval in step 2 include:

[0057] Step 2.1, Local Knowledge Retrieval: Using LLM, extract the entity-level keyword set Kq from query Q. Through semantic embedding model, perform cosine similarity matching between Kq and nodes in knowledge graph G, retrieve a predefined number N of relevant nodes, and obtain the local knowledge set Vq.

[0058] Step 2.2, Global Knowledge Retrieval: Extract global keywords for query Q, match relationships between global keys in the knowledge graph, collect nodes and edges in the multi-hop subgraph of this relationship, and obtain the global knowledge set;

[0059] Step 2.3, Intermediate Layer Connection Knowledge Retrieval: This includes two parts: graph diffusion mechanism and path discovery.

[0060] Part 1: Graph Diffusion Mechanism: Based on a personalized PageRank algorithm, the diffusion vector P is initialized. (0) Through the iterative formula:

[0061] Update vector, where Let A be the restart probability, and A be the normalized adjacency matrix. The iteration... Subsequently, Top-k entities were selected as locally diffused entities, combined with the entity distance attenuation factor. Calculate the diffusion score;

[0062] The second part involves path discovery: constructing two types of reasoning paths R = R1 + R2, where R1 is the set of shortest paths between local knowledge entities and the head and tail entities of global knowledge relationships, and R2 is the set of paths that satisfy cross-layer relevance. Cross-layer semantic matching path, As the threshold, the formula for calculating cross-layer correlation is:

[0063]

[0064] s and t are global relations r G The head and tail entities, w(r) G ) represents the relation weight, and is selected as follows: ,turn up arrive The shortest path to the endpoint.

[0065] Step 3, Two-dimensional screening: Core knowledge is screened by scoring based on relevance, and redundant information is removed by using a diversity strategy to obtain the screened knowledge set;

[0066] Furthermore, the specific steps of the two-dimensional screening in step 3 include:

[0067] Step 3.1, Relevance Dimension Filtering: A comprehensive score Rel2(k,Q) is calculated by weighted fusion of semantic relevance score and knowledge importance score. Sim(.)+ Importance(.) filters out knowledge units that are below a threshold θ. , All are weights; the semantic relevance scoring formula is:

[0068]

[0069] in, Let Q be the knowledge unit to be evaluated, and K be the query. low K is a local keyword set. high The set of global keywords, λ1+λ2=1, is dynamically set according to task characteristics; knowledge importance scoring is designed separately for entities, relationships, and document fragments.

[0070] Entity importance:

[0071]

[0072] Among them, parameters Used to balance semantic and structural features Let be the degree of entity e in the knowledge graph. For entities in a knowledge graph, yes One of the entities, and These are the embedding vectors for entity e and query Q, respectively;

[0073] Importance of the relationship: , For relation embedding vectors, For relationship Weights, parameters It is also used to balance semantic and structural features;

[0074] Importance of document fragments: ;in This is a document fragment. For document fragments The embedding vector;

[0075] Step 3.2, Diversity Dimension Filtering: Based on the maximum boundary relevance strategy, iteratively select the optimal knowledge unit.

[0076]

[0077] in, This is the set of candidate knowledge after initial screening based on relevance. This is a selected set of high-value knowledge, initially empty. The optimal knowledge unit selected for the current round, sim(k, () represents candidate knowledge k and selected knowledge The cosine similarity is used, with parameter γ∈(0,1) as a balance factor. The larger γ is, the more it emphasizes relevance, and the smaller γ is, the more it emphasizes diversity, until a preset number M knowledge units are selected to form the final knowledge set.

[0078] Step 4, Answer Generation: Input the filtered knowledge set as external knowledge into a large language model to generate the target response.

[0079] Furthermore, the specific steps for generating the answer in step 4 include:

[0080] Step 4.1: Using the filtered knowledge set, which includes entity name, relationship type, detailed description and key excerpts from the original text, concatenate them into the input text of "query command + knowledge support + generation requirements";

[0081] Step 4.2: Input text is passed to the target large model. The model analyzes the query requirements and knowledge associations, integrates multi-source information such as entities, relationships, and original text extracts, and establishes a precise mapping between the query and knowledge. Through logical reasoning, the language is organized according to "responding to the core question → supplementing key details" to ensure that the answer is consistent with the query. Figure 1 It provides comprehensive and accurate information, directly generating and outputting the final answer that is adapted to the task scenario.

[0082] The present invention also provides a retrieval enhancement generation system based on multi-level knowledge fusion, the system comprising: a module for executing the retrieval enhancement generation method based on multi-level knowledge fusion.

[0083] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the retrieval enhancement generation method based on multi-level knowledge fusion.

[0084] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the retrieval enhancement generation method based on multi-level knowledge fusion.

[0085] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the retrieval enhancement generation method based on multi-level knowledge fusion.

[0086] To illustrate the effectiveness of the present invention, the following experiments were conducted:

[0087] This experiment aims to verify the effectiveness of a retrieval enhancement generation method based on multi-level knowledge fusion in knowledge-intensive tasks. Specifically, it verifies the role of the intermediate layer connected knowledge (IKL) module in addressing the problem of fragmented knowledge levels, the effectiveness of the two-dimensional filtering framework in filtering noisy information, and the performance differences compared to existing mainstream methods. The embedding models used are GLM-4-Plus and nvidia / NVEmbed-v2, while the generation models are DeepSeek-V3 and GPT-4o-mini. For the data, the Query Focused Summarization (QFS) task uses the Mix, CS, Legal, and Agriculture datasets from the UltraDomain benchmark, containing queries and related documents from different domains; the Multi-Hop Question Answering (MHQA) task randomly samples 1000 multi-hop queries each from 2WikiMultiHopQA and HotpotQA, retaining the original answer labels for evaluation.

[0088] In the experimental design, mainstream methods such as NaiveRAG, GraphRAG, and LightRAG were selected as baselines. For the QFS task, win rate was the core metric, and GPT-4o was used to compare answer quality in pairs across four dimensions: comprehensiveness, practicality, diversity, and overall performance. For the MHQA task, exact matching (EM) and F1 score were used as standard evaluation metrics. All methods had standardized parameter configurations, with core module parameters such as graph diffusion mechanism and two-dimensional screening fixed according to task characteristics.

[0089] Table 1: Win rate comparison with baseline methods in QFS task (%)

[0090] Table 2: EM and F1 scores (%) of the MdRAG method of this invention and the baseline method on the MHQA task.

[0091] Experimental results show that the MdRAG method of this invention significantly outperforms baseline methods in all evaluation dimensions on four major datasets of the QFS task, with an overall win rate of 51.7%-65.2%, demonstrating superior comprehensiveness, practicality, and diversity. In the MHQA task, its exact match (EM) and F1 score both surpass existing state-of-the-art methods, fully demonstrating the synergistic effect of the intermediate layer connecting knowledge modules and the two-dimensional screening framework, effectively improving the accuracy, coherence, and comprehensiveness of answers in knowledge-intensive tasks.

[0092] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A retrieval enhancement generation method based on multi-level knowledge fusion, characterized in that, include: Step 1, Graph Index Construction: The original text document is segmented into fragments, entities and relationships between entities are extracted from the fragments, a knowledge graph with key-value pairs is generated, and the graph operation overhead is reduced by deduplication optimization; Step 2, Multi-level Knowledge Retrieval: Extract keywords from the input query, retrieve local and global knowledge respectively, and then construct semantic associations between local and global knowledge by connecting knowledge modules through an intermediate layer; Step 3, Two-dimensional screening: Core knowledge is screened by scoring based on relevance, and redundant information is removed by using a diversity strategy to obtain the screened knowledge set; Step 4, Answer Generation: Input the filtered knowledge set as external knowledge into a large language model to generate the target response.

2. The retrieval enhancement generation method based on multi-level knowledge fusion according to claim 1, characterized in that: Step 1 includes: Step 1.1, Document Segmentation: Divide the original text document into multiple segments; Step 1.2, Entity and Relationship Extraction: Utilize a large language model to identify entities and relationships between entities in each segment; Step 1.3, Key-value pair generation: Through the parsing function driven by the large language model, key-value pairs are generated for each entity node, with the entity name as the index key and the related text paragraph as the value. For each relation edge, multiple index keys containing the global topic and corresponding text values ​​are generated. Step 1.4, Deduplication Optimization: Merge identical entities and relationships in different segments using a deduplication function to minimize the size of the knowledge graph and ultimately generate the knowledge graph.

3. The retrieval enhancement generation method based on multi-level knowledge fusion according to claim 1, characterized in that: The specific steps of multi-level knowledge retrieval in step 2 include: Step 2.1, Local Knowledge Retrieval: Using LLM, extract the entity-level keyword set Kq from query Q. Through semantic embedding model, perform cosine similarity matching between Kq and nodes in knowledge graph G, retrieve a predefined number N of relevant nodes, and obtain the local knowledge set Vq. Step 2.2, Global Knowledge Retrieval: Extract global keywords for query Q, match relationships between global keys in the knowledge graph, collect nodes and edges in the multi-hop subgraph of this relationship, and obtain the global knowledge set; Step 2.3, Intermediate Layer Connection Knowledge Retrieval: This includes two parts: graph diffusion mechanism and path discovery. Part 1: Graph Diffusion Mechanism: Based on a personalized PageRank algorithm, the diffusion vector P is initialized. (0) Through the iterative formula: ; Update vector, where Let A be the restart probability, and A be the normalized adjacency matrix. The iteration... Subsequently, Top-k entities were selected as locally diffused entities, combined with the entity distance attenuation factor. Calculate the diffusion score; The second part involves path discovery: constructing two types of reasoning paths R = R1 + R2, where R1 is the set of shortest paths between local knowledge entities and the head and tail entities of global knowledge relationships, and R2 is the set of paths that satisfy cross-layer relevance. Cross-layer semantic matching path, As the threshold, the formula for calculating cross-layer correlation is: ； s and t are global relations r G The head and tail entities, w(r) G ) represents the relation weight, and is selected as follows: ,turn up arrive The shortest path to the endpoint.

4. The retrieval enhancement generation method based on multi-level knowledge fusion according to claim 1, characterized in that: The specific steps of the two-dimensional screening in step 3 include: Step 3.1, Relevance Dimension Filtering: A comprehensive score Rel2(k,Q) is calculated by weighted fusion of semantic relevance score and knowledge importance score. Sim(.)+ Importance(.) filters out knowledge units that are below a threshold θ. , All are weights; the semantic relevance scoring formula is: ； in, Let Q be the knowledge unit to be evaluated, and K be the query. low K is a local keyword set. high The set of global keywords, λ1+λ2=1, is dynamically set according to task characteristics; knowledge importance scoring is designed separately for entities, relationships, and document fragments. Entity importance: ； Among them, parameters Used to balance semantic and structural features Let be the degree of entity e in the knowledge graph. For entities in a knowledge graph, yes One of the entities, and These are the embedding vectors for entity e and query Q, respectively; Importance of the relationship: , For relation embedding vectors, For relationship Weights, parameters It is also used to balance semantic and structural features; Importance of document fragments: ;in This is a document fragment. For document fragments The embedding vector; Step 3.2, Diversity Dimension Filtering: Based on the maximum boundary relevance strategy, iteratively select the optimal knowledge unit. ； in, This is the set of candidate knowledge after initial screening based on relevance. This is a selected set of high-value knowledge, initially empty. The optimal knowledge unit selected for the current round, sim(k, () represents candidate knowledge k and selected knowledge The cosine similarity is used, with parameter γ∈(0,1) as a balance factor. The larger γ is, the more it emphasizes relevance, and the smaller γ is, the more it emphasizes diversity, until a preset number M knowledge units are selected to form the final knowledge set.

5. The retrieval enhancement generation method based on multi-level knowledge fusion according to claim 1, characterized in that: The specific steps for generating the answer in step 4 include: Step 4.1: Using the filtered knowledge set, which includes entity name, relationship type, detailed description and key excerpts from the original text, concatenate them into the input text of "query command + knowledge support + generation requirements"; Step 4.2: Input text into the target large model. The model analyzes the query requirements and knowledge associations, integrates multi-source information such as entities, relationships and original text extracts, and establishes a precise mapping between the query and knowledge. Through logical reasoning, the language is organized according to "responding to the core question → supplementing key details" to ensure that the answer is consistent with the query intent, rich in information and accurate in facts, and directly generates and outputs the final answer adapted to the task scenario.

6. A retrieval enhancement generation system based on multi-level knowledge fusion, characterized in that, The system includes a module for executing the retrieval enhancement generation method based on multi-level knowledge fusion as described in any one of claims 1 to 5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the retrieval enhancement generation method based on multi-level knowledge fusion as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the retrieval enhancement generation method based on multi-level knowledge fusion as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the retrieval enhancement generation method based on multi-level knowledge fusion as described in any one of claims 1 to 5.

Images