Mixed retrieval method and system for multi-dimensional heterogeneous knowledge recall enhancement

Abstract

The invention relates to the technical field of information retrieval, and provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method and system.The method comprises the steps that texts recalled through keyword retrieval, sparse vector retrieval and dense vector retrieval are screened through a reciprocal sorting fusion algorithm, and a text type candidate knowledge list is obtained; based on user query, generating and checking a query statement through a large language model, and retrieving an entity-relationship-attribute triple from the knowledge graph library; based on user query, generating enhanced knowledge through a knowledge graph enhanced retrieval method fusing keyword retrieval, vector retrieval and community retrieval; and carrying out format alignment and duplicate removal on the text type candidate knowledge list, the triple result and the enhanced knowledge to form a multi-modal candidate pool, and carrying out reordering score calculation and ordering on each piece of recall knowledge in the multi-modal candidate pool through a reordering model and a business rule to obtain a final retrieval result. And the coverage blind area of single retrieval on heterogeneous knowledge is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information retrieval, and particularly relates to a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Although large models have strong natural language understanding and generation capabilities, in the process of landing in enterprise-level and professional-level application scenarios, large models still face three core difficulties of knowledge timeliness lag, professional field knowledge missing and uncontrollable fact accuracy due to the limitations of technical characteristics. Retrieval augmented generation (RAG) is a key technical solution for large model question and answer landing by supplementing external knowledge and accurately anchoring information. Its core logic is: first, retrieve the authoritative context through the knowledge base, and then inject the retrieval results into the prompt to guide the large model to generate answers based on facts. This mechanism not only makes up for the knowledge gap of the model, but also significantly improves the accuracy and credibility of the answers by introducing verifiable external information sources.

[0004] However, the current mainstream RAG system still has obvious bottlenecks in the retrieval link. The current retrieval method mainly relies on vector retrieval or knowledge graph enhanced retrieval, which cannot fuse multi-dimensional features such as semantic similarity, keyword matching and entity association, and it is difficult to achieve "all-inclusive retrieval". There is a lack of an effective multi-dimensional knowledge fusion retrieval mechanism, which has become a core pain point restricting performance improvement. SUMMARY

[0005] In order to solve the technical problems existing in the background art, the present application provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method and system, which covers unstructured text, semi-structured data and structured graph data through five-way retrieval, solves the coverage blind area of single retrieval for heterogeneous knowledge, and the knowledge graph query and knowledge graph enhanced retrieval provide semantic anchor points for the first three text retrievals, reduce "semantic deviation" and "relationship error", and the multi-path score fusion reduces the misjudgment risk of single retrieval.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions: The first aspect of the present application provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method, which comprises: Obtaining a user query, extracting an entity relationship list, a keyword list and a sentence-level semantic text through a large language model; Based on the entity relationship list and the keyword list, retrieval is performed in the text library through inverted indexing; after the entity relationship list and the keyword list are converted into sparse vectors, retrieval is performed in the sparse vector library through similarity calculation; after the sentence-level semantic text is converted into dense vectors, retrieval is performed in the dense vector library through similarity calculation; the texts recalled by keyword retrieval, sparse vector retrieval and dense vector retrieval are screened through a reverse ranking fusion algorithm to obtain a text type candidate knowledge list; based on the user query, a query sentence is generated and verified through a large language model, and entity-relation-property triples are retrieved from a knowledge graph library; based on the user query, an enhanced knowledge graph retrieval method is used to generate enhanced knowledge by fusing keyword retrieval, vector retrieval and community retrieval. The text type candidate knowledge list, the triple result and the enhanced knowledge are aligned in format and de-duplicated to form a multi-modal candidate pool, and each recalled knowledge in the multi-modal candidate pool is calculated and sorted by a re-ranking model and business rules to obtain the final retrieval result.

[0007] Further, it further comprises: evaluating whether the final retrieval result is sufficient to answer the user query through a large language model, and if not, triggering a new round of retrieval until the information sufficiency threshold is met or the maximum number of iterations is reached.

[0008] Further, the re-ranking score is: ; wherein, represents the re-ranking score of the recalled knowledge i; represents the score output by the re-ranking model; represents the basic fusion score; represents the PageRank score; , and represent weight parameters.

[0009] Further, the basic fusion score is: ; wherein, represents the basic fusion score of the recalled knowledge i; represents a standardization function; , and represent the scores of the recalled knowledge i in the text type candidate knowledge list, the triple result and the enhanced knowledge, respectively; , and represent source importance weights.

[0010] Further, the steps of the knowledge graph enhanced retrieval method comprise: Synchronously starting fusion keyword search, vector search and community search, and performing multi-dimensional search on the multi-dimensional heterogeneous knowledge base; For the multi-dimensional search result, a plurality of candidate nodes are screened through node similarity calculation; The potential association between the candidate nodes is mined through path search and neighbor relationship expansion to obtain candidate relationships; The candidate nodes, candidate relationships and community node search results are integrated to obtain enhanced knowledge.

[0011] Further, the node similarity is: ; Wherein, Q represents a user query, and D represents a multi-dimensional search result.

[0012] Further, the step of generating and verifying the query statement by the large language model comprises: All patterns are taken out from the knowledge graph base, and then matched with the user query to select the best sub-pattern; The user query is input into the query example library to obtain a query example; The query example and the sub-pattern are spliced into a prompt, which is input into the large model to obtain an initial query statement; The initial query statement is input into the knowledge graph base to view the execution plan, if there is an error, the error information is collected, and the syntax verification is continued, in the syntax verification link, the initial query statement is verified again by the large model, if the verification is passed, the query is executed in the knowledge graph base, if the verification fails, the error information is collected, and the modified query statement is generated by inputting the error information into the large model together with the previous error information.

[0013] The second aspect of the application provides a hybrid search system for multi-dimensional heterogeneous knowledge recall enhancement, comprising: An information extraction module configured to: obtain a user query, extract an entity relationship list, a keyword list and a sentence-level semantic text by a large language model; The parallel retrieval module is configured to: based on the entity relationship list and the keyword list, retrieve in a text library through an inverted index; after converting the entity relationship list and the keyword list into sparse vectors, retrieve in a sparse vector library through similarity calculation; after converting the sentence-level semantic text into dense vectors, retrieve in a dense vector library through similarity calculation; through a reverse ranking fusion algorithm, screen the texts recalled by keyword retrieval, sparse vector retrieval and dense vector retrieval to obtain a text type candidate knowledge list; based on a user query, generate and verify a query statement through a large language model, retrieve entity-relation-property triples from a knowledge graph library; based on the user query, through a knowledge graph enhanced retrieval method fusing keyword retrieval, vector retrieval and community retrieval, generate enhanced knowledge; The retrieval fusion module is configured to: align the text type candidate knowledge list, the triple result and the enhanced knowledge in format and remove duplicates to form a multi-modal candidate pool, through a reordering model and business rules, calculate and sort reordering scores of each recalled knowledge in the multi-modal candidate pool to obtain a final retrieval result.

[0014] The third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method described above.

[0015] The fourth aspect of the present application provides a computer device comprising a computer readable storage medium, a processor and a computer program stored on the computer readable storage medium and executable on the processor, the processor executing the program to implement the steps of the multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method described above.

[0016] Compared with the prior art, the present application has the following advantages: The present application covers unstructured text, semi-structured data and structured graph data through five-way retrieval, solves the coverage blind area of single retrieval for heterogeneous knowledge, and the knowledge graph query and knowledge graph enhanced retrieval provide semantic anchors for the first three text retrievals, reduce "semantic deviation" and "relationship error", and the multi-path score fusion reduces the misjudgment risk of single retrieval.

[0017] The dynamic weight distribution of the present application can adapt to different intents such as "exact query", "fuzzy query" and "entity relationship query", and meet the diversified scene needs of enterprise-level intelligent question answering, medical literature retrieval, government policy query and the like. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0019] Figure 1 is a flow chart of a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method according to an embodiment of the application; Figure 2 is a flow chart of the first three retrieval methods according to an embodiment of the application; Figure 3 is a flow chart of the fourth retrieval method according to an embodiment of the application; Figure 4 is a flow chart of the fifth retrieval method according to an embodiment of the application; Figure 5 is a structural schematic diagram of a computer device according to an embodiment of the application. DETAILED DESCRIPTION

[0020] In order to make the objects, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0021] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the application belongs.

[0022] Embodiment one The embodiment provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method.

[0023] Among them, the multi-dimensional heterogeneous knowledge includes unstructured text, semi-structured data, and structured knowledge graph data.

[0024] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment is especially suitable for large model retrieval enhancement generation, enterprise-level intelligent question answering, government knowledge service, and other scenarios that require high recall rate and high precision knowledge matching.

[0025] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment integrates vector semantics, knowledge graph relationship chains, entity attributes, and other heterogeneous features to achieve more accurate and comprehensive knowledge recall, thereby further releasing the potential of RAG in vertical scenarios.

[0026] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment constructs a four-layer hybrid retrieval framework of "reasoning-five-way retrieval parallel-retrieval result fusion-iterative reflection".

[0027] The embodiment provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method, as shown in the formula: Figure 1 The method comprises the following steps: Step 1, reasoning. Reasoning on the user query, and splitting the user query into multiple sub-queries to provide adaptive input for five-way retrieval.

[0028] Specifically, the following operations are performed on the natural language query (denoted as Q) input by the user: (1) Input text preprocessing: Lowercase: unify the text to lowercase form; Full angle to half angle: convert full angle characters to half angle characters; Simplified Chinese: convert traditional Chinese to simplified Chinese; Remove special characters: remove special characters such as punctuation marks, line breaks, etc. using regular expressions; Remove interrogative words: remove common interrogative words (such as “what”, “how” and the like).

[0029] (2) Intention recognition and problem splitting: determine the retrieval intention of Q through a large language model, and split Q into multiple sub-questions to generate adaptive input for five-way retrieval.

[0030] Suppose there is a set scene: User query Q: “What is the outpatient reimbursement ratio of Jinan City in 2025 for employee medical insurance?”; The multi-dimensional heterogeneous knowledge base contains: Text library: 2025 Jinan City Employee Medical Insurance Policy Interpretation PDF document; Sparse vector library: BM25 vector of the above document; Dense vector library: Embedding vector of the above document; Knowledge graph library: contains triples such as “Jinan City-Policy Type-Employee Medical Insurance”, “Employee Medical Insurance-2025 Outpatient Reimbursement Ratio-80% (Retired Personnel 85%) ”.

[0031] For the set scene, input the user query Q into the large language model, extract the entity relationship list (entities “Jinan City” and “employee medical insurance”, relationship “2025 outpatient reimbursement ratio”), generate the keyword list And get the sentence-level semantic text ="Jinan City 2025 Employee Medical Insurance Outpatient Reimbursement Ratio".

[0032] Step 2, five-way retrieval in parallel. Synchronously start five-way retrieval, and perform knowledge recall on the preset multi-dimensional heterogeneous knowledge base (including text library, vector library, and knowledge graph library).

[0033] In this embodiment, five search paths are independent and parallel, sharing the same multi-dimensional heterogeneous knowledge base (pre-constructed, containing text library, sparse vector library, dense vector library, knowledge graph library), as shown in Figure 2 The core execution logic of each path is as follows: (1) First path: keyword search, exact match and fuzzy match are performed on the text library based on inverted index.

[0034] Search object: text library in multi-dimensional heterogeneous knowledge base.

[0035] Core technical means: based on inverted index, the "keyword list + entity relationship list" obtained by preprocessing is used as the search term to perform exact match and fuzzy match.

[0036] Exact match: search for text fragments in the text library that contain the search term (e.g. "employee medical insurance outpatient reimbursement" matches text containing the phrase).

[0037] Fuzzy match: match similar words in the text library to the search term by edit distance (Levenshtein distance) (e.g. "report" matches "report"), where the Levenshtein formula is: ; Where s=0 means the characters are the same, s=1 means they are different, i represents the first i characters of string A, i.e. string A[1..i], j represents the first j characters of string B, i.e. string B[1..j], represents the i-th character of string A, represents the j-th character of string B; if the edit distance is ≤2, it is determined to be a fuzzy match.

[0038] Output result: matching text fragments, text source (such as document ID, page number), exact / fuzzy match label, word frequency matching score (denoted as S1, value 0-1).

[0039] For the given scenario, recall the text fragment containing "Jinan City 2025 Employee Medical Insurance Outpatient Reimbursement Proportion" in the PDF, the matching score =0.9 (word frequency matching).

[0040] (2) Second path: sparse vector search, convert the query into a sparse vector and calculate the similarity with the sparse vector library.

[0041] Search object: sparse vector library in multi-dimensional heterogeneous knowledge base (constructed by sparse vectors converted from text library by TF-IDF or BM25 or SPLADE algorithm).

[0042] Core technical means: convert the pre-processed "keyword list + entity relationship list" into a sparse vector (vector dimension is the size of the vocabulary, non-zero elements correspond to word frequency or BM25 weight or SPLADE weight), calculate the similarity of the sparse vector with the vector in the sparse vector library.

[0043] BM25 similarity calculation: ; where Q represents the query, D represents the document, is the i-th word of the query Q, is the word frequency in the document D, |D| is the length of the document D, avgdl is the average length of all documents, k1 and b are adjustment parameters, , N is the total number of documents, is the number of documents containing the word . The vector calculated by this formula retains the statistical information of the words in the document and the degree of association between the words and the document, which is used to measure the relevance of the document and the query at the term level.

[0044] Output result: Top-K similar text fragments (K can be configured, default K=20), text source, similarity score (denoted as S2, value 0-1).

[0045] For the given scenario, output the above text fragment with Top1 similarity, score =0.85.

[0046] (3) The third way: dense vector retrieval, after vectorizing the query semantics, calculate the cosine similarity with the dense vector library. Specifically, use the embedding model to map the query and the knowledge base text into dense vectors respectively, and get Top-K results and score S3 through cosine similarity calculation.

[0047] Retrieval object: dense vector library in multi-dimensional heterogeneous knowledge base (constructed by dense vectors converted from text library through Embedding model, vector dimension is determined by specific Embeding model, usually 768 / 1024 dimensions).

[0048] Core technical means: input the pre-processed "sentence-level semantic text" (i.e. query Q sem ) into the dense vector model to generate query vector vector_query; calculate the cosine similarity of vector_query with the vector in the dense vector library, and select the text fragment with the highest similarity.

[0049] Cosine similarity calculation: ; where is the query vector, is the document vector.

[0050] Output result: Top-K text snippets, text sources, cosine similarity scores (denoted as S3, taking values 0-1).

[0051] For the given scenario, output the above text snippet with the highest similarity score Top1, score = 0.92.

[0052] (4) Fusion ranking.

[0053] Objective: Fusion ranking of texts retrieved by keyword search, sparse vector search, and dense vector search.

[0054] Core technical means: Through the Reciprocal Rank Fusion (RRF) algorithm, each document in the result list of each route recall is assigned a score according to its ranking position. Generally, the score is the reciprocal of its ranking. For example, the first-ranked document scores 1, the second-ranked document scores 0.5, the third-ranked document scores 0.33, and so on. Therefore, the final score of a document is the sum of its scores in each route recall result.

[0055] RRF score calculation: ; where R = 3 (keyword, sparse vector, and dense vector three-way search), k = 60, is the ranking of document d in the rth route.

[0056] Output result: Top-K text snippets, text sources, RRF scores (denoted as S3-rank, taking values 0-1).

[0057] For the given scenario, output the text snippet with the highest RRF score Top1, score = 0.94.

[0058] (5) Fourth route: Knowledge graph search, based on large language model to generate and verify Cypher query statements, retrieve entity-relation-property triples from knowledge graph database.

[0059] Search object: Knowledge graph database in multi-dimensional heterogeneous knowledge base (including entity nodes, relationship edges, and attribute information, such as "Jinan City - belongs to - Shandong Province", "Employee Medical Insurance - reimbursement ratio - 85%").

[0060] Core technology means: Cypher query generation and verification method based on large model driving, through the fusion of large model semantic understanding, Cypher syntax verification and error regeneration mechanism, the efficient and accurate retrieval of knowledge graph is realized.

[0061] As Figure 3 shown, first, all the schemas are taken out from the graph database, and then matched with Q to select the best sub-schema (sub-schema); Then input the question into the Cypher query example library to get the approximate query example example, splice example and sub-schema into prompt prompt, input into the large model to get the initial Cypher query statement gql; Input gql into the graph database to view the Cypher execution plan, if there is an error error-1, collect the error information, continue to execute the next step Cypher syntax verification; In the Cypher syntax verification link, use the large model to check gql again, if the verification is passed, execute the query in the graph database to get the graph information, if the verification fails, collect the error information error-2, input into the large model together with error-1 to generate the corrected Cypher statement, form the feedback optimization mechanism, and finally according to the confidence score, filter the final knowledge.

[0062] Among them, the confidence score calculation is: ; Among them, represents the weight, which can be adjusted according to business needs, represents the authority of data (such as official policy documents are recorded as 1), is the time decay factor, is the current time, is the data update time, =0.1.

[0063] Output result: matched entity-relation-attribute triple, graph node ID, attribute confidence score (recorded as S4, value 0-1, based on the update time and authority of the graph data) based on data authority and update time.

[0064] For the set scene, the Cypher query gets the triple "employee medical insurance-2025 outpatient reimbursement ratio-85% (retirees 90%) ", the score =0.98.

[0065] (6) Fifth way: knowledge graph enhanced retrieval, combining keywords, vectors and community information to perform graph semantic enhancement recall on the text library, sparse vector library and dense vector library.

[0066] Retrieval object: text library, sparse vector library, dense vector library (node, relationship, community vectorization formed by knowledge graph construction) in multi-dimensional heterogeneous knowledge base.

[0067] Core technical means: knowledge graph enhanced retrieval method integrating keyword retrieval, vector retrieval and community retrieval, suitable for large model retrieval enhancement generation (RAG), intelligent question answering, vertical field knowledge service and other scenarios requiring knowledge graph semantic enhancement and multi-dimensional recall, such as Figure 4 As shown, specifically including: (a) keyword extraction and vectorization processing on user input natural language question; (b) simultaneously starting keyword matching node, vector matching node, vector matching relationship, community node retrieval four paths, and performing multi-dimensional retrieval on the nodes, relationships and communities constructed in the multi-dimensional heterogeneous knowledge base; (c) filtering candidate nodes based on node similarity calculation results; wherein the node filtering calculation formula is as follows: ; Wherein Q represents the query vector and D represents the knowledge vector; (d) mining potential associations between candidate nodes through path search and neighbor relationship expansion to obtain candidate relationships; (e) integrating the filtered candidate node set, candidate relationship set and community node retrieval result to query triplets (i.e. knowledge representation units such as "entity-relation-entity", "entity-attribute-value"); (f) returning the final knowledge graph enhanced retrieval result, enhanced knowledge.

[0068] Output result: enhanced text segment, graph anchoring relationship (such as "Text Segment 1 - Anchoring Entity: Jinan City"), enhanced adaptation score (denoted as S5, value 0-1, calculated based on retrieval word expansion matching degree and relationship conflict rate).

[0069] For the given scenario, the output result: the knowledge graph contains nodes "Jinan City", "employee medical insurance", "outpatient reimbursement ratio", relationship "contains", attribute "outpatient reimbursement ratio-2025 value-first class hospital 80%", "outpatient reimbursement ratio-2025 value-second class hospital 70%", "outpatient reimbursement ratio-2025 value-third class hospital 60%", recall the above text segment, score =0.93.

[0070] Step 3, retrieval result fusion. Align, deduplicate and reorder the intermediate results of the five-way retrieval, and score and sort each recalled knowledge by introducing a reordering model and business rules.

[0071] The output results of the five-way search are subjected to a three-step fusion of "alignment, deduplication, and reordering", as follows: The text-based candidate knowledge list is aligned in format and deduplicated with the triple results of knowledge graph search and the enhanced knowledge (such as text + graph relationship with semantic anchor points) of graph-enhanced search, forming a multi-modal candidate pool containing text knowledge, graph knowledge, and enhanced knowledge. Ranking output: The similarity between the problem and the recalled knowledge in the multi-modal candidate pool is calculated using the reordering module, and the final ranking score of each candidate knowledge is output. The top-N (N is configurable, and the default N = 30) results are selected as the final recall list, which is the final search result.

[0072] The specific steps are as follows: (1) Unify the formats of text segments, triples, and enhanced knowledge to construct a multi-modal candidate pool. (2) Receive multiple deduplicated candidate results and their original scores from keyword and vector hybrid search, knowledge graph search, and knowledge graph relationship-enhanced search. (3) Through a first set of weight parameters , the normalized scores of the candidate results from different sources are linearly weighted to obtain a basic fusion score . (4) Through a pre-trained reordering model, a reordering score is calculated for each candidate result. (5) Through a second set of weight parameters , the reordering score , the basic fusion score , and the PageRank score are fused to obtain a final score . (6) All candidate results are uniformly sorted according to the final score .

[0073] The first set of weight parameters and the second set of weight parameters can be dynamically adjusted by administrators through a configuration interface to adapt to different business scenario strategies.

[0074] Core technical means: The basic fusion score calculation formula is as follows: ; wherein, represents the basic fusion score of knowledge i; represents a standardization function that maps scores from different sources independently to the [0, 1] interval; , , These represent the original scores of knowledge i in fusion ranking, knowledge graph retrieval, and knowledge graph-enhanced retrieval, respectively. , , Indicates the importance weight of the source, and The weight can be dynamically adjusted according to business needs.

[0075] By introducing a re-ranking model and business rules, the final re-ranking score is calculated using the following formula: ; in, Represents the final reordering score of knowledge i; This represents the score output by the reordering model, with a value range of [0,1]. Indicates the basic integration score; This represents the PageRank score; , , The weighting parameters representing the final score are the most important business control settings. If the re-ranking model is to dominate the scores, then set... A higher value, such as 0.7, can be set if the authority of the knowledge source is required. It is a relatively high value; in, , This represents the PageRank score of the i-th knowledge, and the sum of the PageRanks of all knowledge is 1; d is the damping coefficient, which defaults to 0.85; N represents the total number of knowledge. This represents the set of all knowledge points pointing to knowledge i; This represents the edge weight from knowledge j to knowledge i; This represents the set of out-chain neighbors of knowledge j, that is, all the knowledge that j points to; This represents the sum of the weights of all outgoing links to knowledge j. PageRank itself operates on the link relationships between nodes. Now, each search result is treated as a graph node, and edges are established according to the logic that "if result A and result B are highly related (semantic / entity / topic consistency), then A→B and B→A (or one-way) can be regarded as mutual recommendation." In this way, PageRank can automatically identify the most central and authoritative result in the overall result set through the mechanism of "related results voting for each other." In this embodiment, after constructing all deduplicated search results (from hybrid search, knowledge graph search, and enhanced search) into a graph, the "global importance score" of each result is calculated, thereby realizing the requirement of multi-source search fusion.

[0076] Step 4, Iterative reflection. Reflect on the results of the fusion, judge whether to adjust the scheme, and perform a new round of query iteration, finally output an ordered knowledge recall list, and provide structured knowledge support for downstream applications (such as RAG, intelligent question answering).

[0077] Specifically, through the large language model, it is evaluated whether the current recalled knowledge is sufficient to answer the user query. If not, the subquery strategy is adjusted based on the reflection result (the large model generates multiple query statements according to the user's query and pre-defined business knowledge, that is, the large model generates multiple queries), and a new round of five-way retrieval is triggered, until the information sufficiency threshold is met or the maximum iteration number is reached, and finally the structured ordered knowledge recall list is output.

[0078] Specifically, it includes: (1) Reflection and refinement: review the retrieved information through the large model, reflect on whether the current information is sufficient to support answering the user's query, and adjust the scheme as needed; (2) Iteration and output: multiple iterations of results to obtain reliable and detailed knowledge, and then output the final recall list after fusion in a structured format, which includes: Knowledge content (text segment or entity-relation-attribute triple); Knowledge source (document ID, graph node ID, data update time); Knowledge score (comprehensive score S_total); Knowledge type label (“text segment”, “knowledge graph triple”, “community report”).

[0079] For the set scene, the final recall list Top1 is “knowledge content: The medical insurance reimbursement rate for outpatients in Jinan City in 2025 is 80%, and for retirees is 85%; source: “Jinan City Medical Insurance Policy Interpretation” P5, graph node ID: medical insurance_2025_01; comprehensive score: 0.95; knowledge type: unstructured text + structured graph data”.

[0080] As can be seen from the above embodiments, the present application can effectively integrate the advantages of multi-way retrieval, realize accurate recall of heterogeneous knowledge, and provide reliable knowledge support for intelligent question answering.

[0081] The embodiment provides a multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method, which aims to solve the problems of low recall rate and insufficient precision caused by a single retrieval path in the professional scene of an existing retrieval augmentation generation (RAG) system. A four-layer architecture of "reasoning-five-way retrieval parallel-result fusion-iterative reflection" is constructed: first, the reasoning module is used for standardized preprocessing and intent recognition of the user query; then, five-way heterogeneous retrieval paths are executed in parallel, including keyword retrieval, sparse vector retrieval, dense vector retrieval, knowledge graph retrieval, and knowledge graph enhanced multi-dimensional retrieval; then, the fusion module is used for format alignment, deduplication and reordering of the multi-dimensional results; finally, an iterative reflection mechanism driven by a large language model is introduced to dynamically evaluate the recall quality and trigger optimization iteration. The multi-dimensional features such as semantics, keywords, entity relationships and community structure are effectively fused, and the comprehensiveness and accuracy of knowledge recall are significantly improved, which is suitable for high-demand application scenarios such as RAG, enterprise intelligent question answering and government knowledge service.

[0082] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment improves the comprehensiveness of knowledge recall: through five-way retrieval, unstructured text, semi-structured data and structured graph data are covered, the coverage blind spot of single retrieval for heterogeneous knowledge is solved, and "all-inclusive retrieval" is realized.

[0083] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment optimizes the matching precision: the knowledge graph query and the knowledge graph enhanced retrieval provide semantic anchors for the first three text retrievals, reduce "semantic deviation" and "relationship error", and the multi-path score fusion reduces the misjudgment risk of single retrieval.

[0084] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment has strong scene adaptability: dynamic weight allocation can adapt to different intents such as "precise query", "fuzzy query" and "entity relationship query", and meet the diversified scene needs of enterprise-level intelligent question answering, medical literature retrieval and government policy query.

[0085] The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method provided by the embodiment enhances explainability: the knowledge graph retrieval and enhanced retrieval provide entity relationship and source annotation, solve the "black box" problem of traditional vector retrieval, provide traceable knowledge basis for downstream large model generation, and reduce the "illusion" phenomenon.

[0086] Embodiment two The multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval system provided by the embodiment comprises: The information extraction module is configured to: acquire a user query, extract an entity relationship list, a keyword list and a sentence-level semantic text through a large language model; The parallel retrieval module is configured to: perform retrieval in a text database using an inverted index based on an entity relation list and a keyword list; convert the entity relation list and keyword list into sparse vectors and then perform retrieval in a sparse vector database using similarity calculation; convert sentence-level semantic text into dense vectors and then perform retrieval in a dense vector database using similarity calculation; filter the texts recalled by keyword retrieval, sparse vector retrieval, and dense vector retrieval using an inverse sorting fusion algorithm to obtain a candidate knowledge list for text classes; generate and validate query statements based on user queries using a large language model, and retrieve entity-relationship-attribute triples from a knowledge graph database; and generate enhanced knowledge by integrating keyword retrieval, vector retrieval, and community retrieval knowledge graph enhancement methods based on user queries. The retrieval fusion module is configured to: perform format alignment and deduplication on the text-based candidate knowledge list, triple results, and enhanced knowledge to form a multimodal candidate pool; and calculate and sort the re-ranking score for each recalled knowledge in the multimodal candidate pool using a re-ranking model and business rules to obtain the final retrieval results.

[0087] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0088] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the hybrid retrieval method for enhancing multidimensional heterogeneous knowledge recall as described in Embodiment 1 above.

[0089] Example 4 This embodiment provides a computer device, such as... Figure 5 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps of the hybrid retrieval method for enhanced multidimensional heterogeneous knowledge retrieval as described in Embodiment 1 above.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement, characterized in that, Comprise: Obtain a user query, extract an entity relationship list, a keyword list, and a sentence-level semantic text through a large language model; Based on the entity relationship list and the keyword list, retrieve in the text library through inverted indexing; convert the entity relationship list and the keyword list into sparse vectors, and retrieve in the sparse vector library through similarity calculation; convert the sentence-level semantic text into a dense vector, and retrieve in the dense vector library through similarity calculation; filter the texts recalled by keyword retrieval, sparse vector retrieval, and dense vector retrieval through a reverse ranking fusion algorithm to obtain a text type candidate knowledge list; based on the user query, generate and verify a query statement through a large language model, and retrieve entity-relation-property triples from a knowledge graph library; Based on the user query, generate enhanced knowledge through a knowledge graph enhanced retrieval method that fuses keyword retrieval, vector retrieval, and community retrieval; Align and deduplicate the text type candidate knowledge list, the triple result, and the enhanced knowledge to form a multi-modal candidate pool, reorder each recalled knowledge in the multi-modal candidate pool through a reordering model and business rules to calculate a reordering score and sort to obtain the final retrieval result.

2. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement as claimed in claim 1, wherein, Also include: Evaluate whether the final retrieval result is sufficient to answer the user query through a large language model, if not, trigger a new round of retrieval until the information sufficiency threshold is met or the maximum number of iterations is reached.

3. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement as claimed in claim 1, wherein, The reordering score is: ; wherein, represents the reordering score of recall knowledge i; represents the score output by the reordering model; represents the base fusion score; represents the PageRank score; , and represent weight parameters.

4. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement as claimed in claim 3, wherein, The base fusion score is: ; wherein, denotes the base fusion score for recall knowledge i; denotes a normalization function; , and denote the score of recall knowledge i in the list of text-based candidate knowledge, triple results and enhanced knowledge, respectively; , and denote the source importance weight.

5. The hybrid search method of claim 1, wherein, The steps of the knowledge graph enhanced retrieval method include: Simultaneously start keyword retrieval, vector retrieval, and community retrieval to perform multi-dimensional retrieval on a multi-dimensional heterogeneous knowledge base; For multi-dimensional retrieval results, filter a number of candidate nodes through node similarity calculation; Dig up potential associations between candidate nodes through path search and neighbor relationship expansion to obtain candidate relationships; Integrate candidate nodes, candidate relationships, and community node retrieval results to obtain enhanced knowledge.

6. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement as claimed in claim 1, wherein, The node similarity is: ； Wherein, Q represents a user query, and D represents multi-dimensional retrieval results.

7. A hybrid search method with multi-dimensional heterogeneous knowledge recall enhancement as claimed in claim 1, wherein, The steps of generating and verifying a query statement through a large language model include: Take all patterns from the knowledge graph library, then match them with the user query, and select the best sub-pattern; Input the user query into the query example library to obtain a query example; Concatenate the query example and the sub-pattern as a prompt, and input it into the large model to obtain an initial query statement; Input the initial query statement into the knowledge graph library to view the execution plan, if there is an error, collect the error information, continue to perform syntax verification, and use the large model to verify the initial query statement again in the syntax verification link, if the verification is passed, execute the query in the knowledge graph library, if the verification fails, collect the error information, and input it into the large model together with the previous error information to generate a corrected query statement.

8. A hybrid search system with multi-dimensional heterogeneous knowledge recall enhancement, characterized in that, Include: An information extraction module configured to: obtain a user query, extract an entity relationship list, a keyword list, and a sentence-level semantic text through a large language model; The parallel retrieval module is configured to: based on the entity relationship list and the keyword list, retrieve in the text library through an inverted index; after converting the entity relationship list and the keyword list into sparse vectors, retrieve in a sparse vector library through similarity calculation; after converting the sentence-level semantic text into a dense vector, retrieve in a dense vector library through similarity calculation; for the texts recalled by keyword retrieval, sparse vector retrieval and dense vector retrieval, screen through a reverse ranking fusion algorithm to obtain a text type candidate knowledge list; based on a user query, generate and verify a query statement through a large language model, and retrieve entity-relation-property triples from a knowledge graph library; Based on the user query, an enhanced knowledge graph retrieval method is generated by fusing keyword retrieval, vector retrieval and community retrieval, and an enhanced knowledge is generated; The retrieval fusion module is configured to: align and deduplicate the text type candidate knowledge list, the triple result and the enhanced knowledge to form a multi-modal candidate pool, calculate and sort the reordering scores of each recalled knowledge in the multi-modal candidate pool through a reordering model and business rules to obtain the final retrieval result.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method of any one of claims 1-7.

10. A computer device, comprising a computer readable storage medium, a processor, and a computer program stored on the computer readable storage medium and executable on the processor, wherein, The processor executes the program to implement the steps in the multi-dimensional heterogeneous knowledge recall enhanced hybrid retrieval method of any one of claims 1-7.

Images