Multi-source information knowledge fusion and intelligent retrieval methods and systems

Abstract

This invention relates to a method and system for multi-source information knowledge fusion and intelligent retrieval. The method includes: acquiring multi-source heterogeneous data and performing unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework. Based on the unified knowledge framework, a large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples. The structured knowledge triples are then structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body. User query requests are received and responded to by performing deep analysis on the user input data to obtain the user's query intent and query elements. A condition graph is constructed based on the user's query intent and query elements, and the condition graph is mapped to a knowledge network for entity localization and condition graph matching to generate structured retrieval results. This reduces network overhead and scheduling latency, and improves retrieval response speed and accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for multi-source information knowledge fusion and intelligent retrieval. Background Technology

[0002] As digital transformation deepens, the data accumulated across industries is characterized by being multi-source, heterogeneous, and massive. This data includes structured databases, unstructured documents, semi-structured tables, and multimodal content such as images and audio. Due to differences in their sources, formats, and semantic systems, these data form isolated information silos, making it difficult to effectively link, understand, and quickly retrieve key knowledge. This, to some extent, restricts the efficiency of intelligent decision-making and knowledge services.

[0003] Currently, to address the aforementioned technical challenges, existing technologies are evolving primarily in three directions. First, by constructing graph structures of entities, relationships, and attributes, a semantic association framework for data is provided. This approach excels at handling structured queries and multi-hop reasoning, but its depth and flexibility in semantic understanding of unstructured text are relatively limited. Second, large language models, with their strong natural language understanding and generation capabilities, handle open-domain question answering. For example, RAG technology enhances the accuracy and timeliness of generated answers by retrieving relevant document fragments. However, this approach typically relies on vector similarity retrieval and lacks the ability to explicitly model and reason about the inherent relationships within structured knowledge. Third, to balance structured queries and semantic search, systems integrating multiple retrieval modes have been proposed. These systems attempt to cover a wider range of query needs by calling multiple retrieval engines in parallel or sequentially and fusing the retrieval results.

[0004] Therefore, existing technologies require the maintenance of multiple independent engines at runtime, and the routing decision logic is complex and prone to errors. Each query may involve multiple network calls and calculations, leading to accumulated latency. In addition, the results from different engines are at different semantic levels, making it difficult to reconstruct the semantic relationships between the results of different engines through subsequent information fusion, which in turn leads to information bias or loss. Summary of the Invention

[0005] The present invention aims to provide a method and system for multi-source information knowledge fusion and intelligent retrieval to address the shortcomings of existing technologies. The technical problem to be solved by the present invention is achieved through the following technical solutions.

[0006] This invention provides a method for multi-source information knowledge fusion and intelligent retrieval, the method comprising:

[0007] Acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework;

[0008] Based on the unified knowledge framework, the large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples.

[0009] The structured knowledge triples are structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body.

[0010] Receive user query requests and respond to the user query requests by performing in-depth analysis of user input data to obtain user query intent and query elements;

[0011] A condition graph is constructed based on the user's query intent and query elements, and the condition graph is mapped to a knowledge network for entity localization and condition graph matching to generate structured search results.

[0012] In one embodiment, the step of acquiring multi-source heterogeneous data and performing unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework includes:

[0013] Knowledge domain analysis is performed on the multi-source heterogeneous data to construct a hierarchical domain ontology based on the multi-source heterogeneous data. The domain ontology includes entity types, entity attributes, and entity relationships of the target knowledge domain.

[0014] The domain ontology is transformed into a knowledge graph containing entity types and corresponding entity attributes and entity relationships for each entity type. The multi-source heterogeneous data is then standardized based on the knowledge graph to construct the unified knowledge framework.

[0015] In one embodiment, the step of invoking a large language model to perform structured preprocessing on the multi-source heterogeneous data based on the unified knowledge framework to generate structured knowledge triples includes:

[0016] Under the unified knowledge framework, entity recognition, entity relation extraction and entity attribute extraction are performed on the text data in the multi-source heterogeneous data through the large language model, and the knowledge domain corresponding to the current entity is written into the model prompt words to generate the structured knowledge triples.

[0017] The tabular data in the multi-source heterogeneous data is transformed into tabular description text through a large language model, and the tabular description text is mapped to the target entities and entity attributes in the knowledge domain.

[0018] In one embodiment, the step of invoking a large language model to perform structured preprocessing on the multi-source heterogeneous data based on the unified knowledge framework to generate structured knowledge triples further includes:

[0019] Text data is extracted from the image data in the multi-source heterogeneous data using OCR or a parsing library, and the image description text is parsed. The text data and image description text in the image data are then input into a large language model to generate the structured knowledge triples.

[0020] The structured knowledge triples are subjected to entity implicit relation reasoning and knowledge completion processing using a large language model, and conflict detection is performed on the structured knowledge triples to resolve the detected conflicts by invoking a preset strategy.

[0021] In one embodiment, the step of structuring the structured knowledge triples using an attribute graph and vectorizing the associated data of each entity using an embedding model to obtain a fused knowledge body includes:

[0022] The original documents are extracted from the multi-source heterogeneous data, and the semantic boundaries of the original documents are identified by natural language processing technology. The original documents are then sliced ​​and metadata is annotated according to the semantic boundaries to obtain multiple document slices carrying metadata.

[0023] Extract the target entity from the document slice and establish a bidirectional index between the target entity and the fused knowledge body, so as to set semantic tags for the document slice according to the bidirectional index.

[0024] In one embodiment, the step of structuring the structured knowledge triples using an attribute graph and vectorizing the associated data of each entity using an embedding model to obtain a fused knowledge body further includes:

[0025] Each entity in the structured knowledge triple and the entity relationships between different entities are transformed into an attribute graph, and the attribute graph is stored in a graph database.

[0026] The entity description text, entity attributes, and document slices of each entity in the structured knowledge triple are transformed into high-dimensional vectors by an embedding model, and the high-dimensional vectors are stored in a vector database.

[0027] In one embodiment, the step of constructing a condition graph based on the user's query intent and query elements, and mapping the condition graph to a knowledge network for entity localization and condition graph matching to generate structured search results includes:

[0028] The conditional graph is constructed using the user's query intent and the query entities in the query elements as nodes, the entity relationships between different query entities as edges, and the entity attributes of the query entities as constraints.

[0029] Calculate the vector similarity between entity nodes in the condition graph and entity nodes in the fused knowledge body, and locate the anchor entity between the condition graph and the fused knowledge body based on the vector similarity;

[0030] Starting from the anchor entity, multiple knowledge subgraphs that match the topological structure and semantic tags of the condition graph are searched in the knowledge network, and the multiple knowledge subgraphs are aggregated and semantically enhanced to obtain the structured retrieval results.

[0031] This invention also provides a multi-source information knowledge fusion and intelligent retrieval system for implementing the multi-source information knowledge fusion and intelligent retrieval method described in any of the above claims, the system comprising:

[0032] The knowledge modeling module is used to acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework.

[0033] The structured processing module is used to perform structured preprocessing on the multi-source heterogeneous data by calling the large language model based on the unified knowledge framework, so as to generate structured knowledge triples.

[0034] The knowledge fusion module is used to perform structuring processing on the structured knowledge triples through attribute graphs, and to vectorize and store the associated data of each entity through an embedding model to obtain a fused knowledge body.

[0035] The query analysis module is used to receive user query requests and, in response to the user query requests, perform in-depth analysis of user input data to obtain user query intent and query elements;

[0036] The intelligent retrieval module is used to construct a condition graph based on the user's query intent and query elements, and to map the condition graph to a knowledge network for entity localization and condition graph matching to generate structured retrieval results.

[0037] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the multi-source information knowledge fusion and intelligent retrieval method as described above.

[0038] The present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the multi-source information knowledge fusion and intelligent retrieval method as described above.

[0039] The aforementioned multi-source information knowledge fusion and intelligent retrieval method and system acquires multi-source heterogeneous data and performs unified knowledge modeling on this data to obtain a unified knowledge framework. Then, based on this unified knowledge framework, a large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples. Next, attribute graphs are used to further structure these knowledge triples, and an embedding model is used to vectorize and store the associated data of each entity, resulting in a fused knowledge body. Subsequently, user query requests are received and responded to by performing deep analysis of the user input data to obtain the user's query intent and query elements. Finally, a condition graph is constructed based on the user's query intent and query elements, and this condition graph is mapped to the knowledge network for entity localization and condition graph matching to generate structured retrieval results. This invention introduces a conditional graph as a unified virtual query interface, directly compiling the user's query intent into a graph structure query. This query can be directly performed on a unified fused knowledge body in the background for one-time graph pattern matching and semantic search, effectively eliminating the network overhead and scheduling delay caused by multi-engine calls in existing technologies. This simplifies the system architecture while improving the retrieval response speed and the accuracy of retrieval results. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the multi-source information knowledge fusion and intelligent retrieval method provided by this invention;

[0042] Figure 2 This is a schematic diagram of the overall system architecture of the multi-source information knowledge fusion and intelligent retrieval method provided in a specific embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the multimodal knowledge fusion process of the multi-source information knowledge fusion and intelligent retrieval method provided in a specific embodiment of the present invention;

[0044] Figure 4 A schematic diagram of the document slicing process of the multi-source information knowledge fusion and intelligent retrieval method in a specific embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of the conditional graph intelligent retrieval process of the multi-source information knowledge fusion and intelligent retrieval method provided in a specific embodiment of the present invention;

[0046] Figure 6A schematic diagram of the structure of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention;

[0047] Figure 7 This is a diagram of the internal structure of the electronic device provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0049] The following is combined Figures 1 to 7 This invention describes a multi-source information knowledge fusion and intelligent retrieval method and system.

[0050] like Figure 1 As shown, in one embodiment, a multi-source information knowledge fusion and intelligent retrieval method includes the following steps:

[0051] Step S110: Obtain multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework.

[0052] Specifically, the system acquires multi-source heterogeneous data from various channels and performs unified knowledge modeling on the multi-source heterogeneous data to construct a unified knowledge framework.

[0053] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention includes the following steps in step S110:

[0054] Step S111: Perform knowledge domain analysis on the multi-source heterogeneous data to construct a hierarchical domain ontology based on the multi-source heterogeneous data. The domain ontology includes entity types, entity attributes, and entity relationships of the target knowledge domain.

[0055] Step S112: The domain ontology is transformed into a knowledge graph containing entity types and the entity attributes and entity relationships corresponding to each entity type. The multi-source heterogeneous data is then standardized based on the knowledge graph to construct a unified knowledge framework.

[0056] Combination Figure 2 and Figure 3As shown in the specific embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention has an overall architecture divided into a data source layer, a data processing and knowledge generation layer, a knowledge storage and management layer, an intelligent engine layer, an intelligent service and interaction layer, and an intelligent application layer. First, the data source layer receives multi-source heterogeneous data from various channels, such as text, images, videos, tables, and databases. Then, the data processing and knowledge generation layer is responsible for unified knowledge modeling and multimodal knowledge fusion. In the unified knowledge modeling process, a unified framework of entities, relationships, and attributes is defined by constructing a hierarchical domain knowledge base, providing standardized semantic tags for all received data. Then, unified knowledge modeling of the multi-source heterogeneous data is integrated, and multimodal knowledge fusion is performed based on a large model. During the multimodal knowledge fusion process, the cross-modal understanding capability of the large language model is used to perform deep semantic parsing, entity alignment, entity relationship extraction, and completion on the original data according to the above framework. After processing by this layer, a unified fused knowledge body is generated, which is a structured, semantically rich, reasonable knowledge network.

[0057] In this embodiment, a unified framework of entities, relationships, and attributes is defined by constructing a hierarchical domain knowledge base, providing standardized semantic tags for all data, including two processing steps: domain ontology construction and knowledge image schema definition.

[0058] In the construction of the domain ontology, core concepts, entity types, attributes, and relationship types are abstracted by analyzing the target business domain, constructing a hierarchical domain ontology to clarify the hierarchical relationships and constraints between entities. During the definition of the knowledge graph schema, the domain ontology is concretized into an operable knowledge graph schema, including entity type and attribute schemas and relationship type schemas. The entity type and attribute schema defines standardized attribute fields and data types for each type of entity. The relationship type schema defines allowed relationship types, the starting and ending entity type constraints of relationships, and the attributes of the relationships themselves. This schema serves as a mandatory specification for all data import, cleaning, and mapping.

[0059] Step S120: Based on the unified knowledge framework, the large language model is called to perform structured preprocessing on multi-source heterogeneous data to generate structured knowledge triples.

[0060] Specifically, the unified knowledge framework obtained by modeling in step S110 calls the large language model to perform structured preprocessing on the acquired multi-source heterogeneous data to generate structured knowledge triples.

[0061] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention includes the following steps in step S120:

[0062] Step S121: Under the unified knowledge framework, entity recognition, entity relation extraction and entity attribute extraction are performed on text data in multi-source heterogeneous data through a large language model, and the knowledge domain corresponding to the current entity is written into the model prompt words to generate structured knowledge triples.

[0063] Step S122: Convert the tabular data in the multi-source heterogeneous data into tabular description text through the large language model, and map the tabular description text to the target entity and entity attributes in the knowledge domain.

[0064] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention further includes the following steps in step S120:

[0065] Step S123: Extract text data from image data in multi-source heterogeneous data using OCR or parsing library, and parse image description text to input text data and image description text from image data into large language model to generate structured knowledge triples.

[0066] Step S124: The structured knowledge triples are subjected to entity implicit relation reasoning and knowledge completion processing through the large language model, and conflict detection is performed on the structured knowledge triples to resolve the detected conflicts by calling a preset strategy.

[0067] Combination Figure 3 As shown in the specific embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention utilizes the cross-modal understanding capability of a large language model. Under the guidance of a unified knowledge framework, the original data is subjected to deep semantic parsing, entity alignment, relation extraction and completion according to the framework, thereby generating structured knowledge triples and vectors. This process includes four processing steps: multi-modal data parsing and semantic extraction, cross-source entity alignment and disambiguation, implicit relation reasoning and knowledge completion, and conflict detection and resolution.

[0068] In the process of multimodal data parsing and semantic extraction, each data unit is transformed into a preliminary set of <entity, relation, entity> triples or <entity, attribute, value> triples with confidence. Different data parsing and semantic extraction methods are used for different data types. For example, text data is directly input into the large model for entity recognition, relation extraction, and attribute extraction. The domain schema is injected into the model prompt to guide the model to output according to the specification. For tabular data types, the table structure (header, rows, columns) is transformed into descriptive text or directly parsed into <header, cell value> pairs. The large model understands the semantics of the table header and maps the understood semantics to entities or attributes in the domain schema. For image or PDF document types, text is extracted through OCR or a dedicated parsing library, and then combined with image description models or layout analysis to obtain semantic information of mixed text and images, which is then given to the large model for comprehensive understanding.

[0069] In the process of cross-source entity alignment and disambiguation, the initial triples from different data sources may contain entities that point to the same object but have different descriptions. The system calls the large model to perform similarity calculation and semantic discrimination based on entity name, context description, and key attributes to identify synonymous entities and entities with the same name but different meanings. Then, an entity alignment mapping table is generated to map multiple entity identifiers with different appearances to a unified and standardized entity ID.

[0070] In the process of implicit relation reasoning and knowledge completion, the large model is invoked to perform logical reasoning based on the extracted entities and explicit entity relationships, combined with the prior knowledge of the large model itself. This can discover and supplement implicit entity relationships that are not explicitly stated in the original data, thereby enriching the density and depth of the knowledge network.

[0071] During the conflict detection and resolution process, information from different sources may conflict. The system detects conflicts by setting rules or models based on dimensions such as the credibility of the information source, timestamp, and consistency with other facts. When a conflict is detected, a strategy is adopted to resolve it to ensure the consistency within the knowledge body.

[0072] Step S130: The structured knowledge triples are processed by attribute graphs, and the associated data of each entity is vectorized and stored by embedding model to obtain the fused knowledge body.

[0073] Specifically, the system stores structured knowledge triples in a graph database in the form of attribute graphs to achieve structured processing of structured knowledge triples. At the same time, it uses an embedding model to vectorize the entity parameters involved in each entity in the structured knowledge triples and stores them in a vector database. Finally, the unified fused knowledge body is a structured, semantically rich, reasonable knowledge network, which is stored and represented by a knowledge graph database and a vector database.

[0074] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention includes the following steps in step S130:

[0075] Step S131: Extract the original document from the multi-source heterogeneous data, and identify the semantic boundary of the original document through natural language processing technology, so as to slice the original document and annotate the metadata according to the semantic boundary, and obtain multiple document slices carrying metadata.

[0076] Step S132: Extract the target entity from the document slice and establish a bidirectional index between the target entity and the fused knowledge body, so as to set semantic tags for the document slice according to the bidirectional index.

[0077] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention further includes the following steps in step S130:

[0078] Step S133: Transform each entity in the structured knowledge triple and the entity relationships between different entities into an attribute graph, and store the attribute graph in the graph database.

[0079] Step S134: The entity description text, entity attributes, and document slices of each entity in the structured knowledge triple are transformed into high-dimensional vectors through the embedding model, and the high-dimensional vectors are stored in the vector database.

[0080] In specific embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by this invention uses a knowledge storage and management layer to design a general knowledge base, a dedicated knowledge base, and a personal knowledge base, based on the degree of knowledge sharing, update frequency, and application scenarios. The general knowledge base stores cross-domain, highly consensus-based, and relatively stable common knowledge; the dedicated knowledge base stores professional knowledge, attributes, rules, processes, and internal data specific to a particular domain, and, based on the general knowledge base, integrates and expands a general ontology to form domain depth; the personal knowledge base stores personalized information closely related to specific users, supporting the second-level construction of personal knowledge views. All three types of knowledge bases adopt a hybrid storage architecture of graph databases and vector databases, respectively used to store structured relationships and semantic vectors. The intelligent engine layer includes a retrieval enhancement generation engine and a knowledge fusion and reasoning engine. The retrieval enhancement generation engine, relying on the RAG engine and intelligent document slicing, performs intelligent segmentation of unstructured long documents based on semantic integrity, establishing bidirectional links between each document slice and relevant entity nodes in the fused knowledge body to form an enhanced index, enabling deep fusion of text fragments and structured knowledge.

[0081] In this embodiment, the high-quality knowledge generated above after cleaning, alignment, completion and resolution is persistently stored as a fused knowledge body that can be efficiently queried, including three processing steps: structured storage, vectorized storage and bidirectional link establishment.

[0082] In the structured storage process, all standardized entities and their explicit and implicit relationships are stored in a graph database in the form of attribute graphs. In the vectorized storage process, a vector database is used. The text descriptions, key attributes, and important semantic slices from the original document for each entity are transformed into high-dimensional vectors through an embedding model and stored in the vector database. The final output is a unified fused knowledge body, which is a structured, semantically rich, reasonable knowledge network, stored and represented using both a knowledge graph and a vector database.

[0083] Combination Figure 4 As shown, in this embodiment, document slicing is used to transform the original long document into independent knowledge units that maintain semantic integrity. This includes two processes: semantic analysis and boundary recognition, and slice encapsulation and metadata annotation. In the semantic analysis and boundary recognition process, natural language processing techniques are used to identify the natural semantic boundaries of the document, including structural boundaries, semantic boundaries, and contextual overlap. In the slice encapsulation and metadata annotation process, rich metadata is appended to each generated document slice to form a structured document block. The metadata includes identification information, location information, and a semantic summary.

[0084] After document slicing, a computable link needs to be established between the unstructured text slices and the structured knowledge network. This involves two processes: entity recognition and linking, and bidirectional indexing. Entity recognition and linking refers to using entity recognition technology to extract entities mentioned in each document slice and linking these mentioned entities with existing standardized entity nodes in the fused knowledge body to establish a bidirectional index. The bidirectional index includes establishing a forward index between the slice and the knowledge subgraph, recording which entity nodes in the knowledge network the document slice is associated with, thus assigning the slice a deep semantic label; and establishing a reverse index between entity nodes and slices, recording the IDs of all document slices that reference or describe the entity and their contextual summaries on the entity nodes in the knowledge network, providing rich original evidence support for the entity nodes.

[0085] Finally, the output includes a document slice library with semantic tags, bidirectional links between document slices and knowledge bodies, including a link table from each slice to a knowledge grid node and a reverse link table from each knowledge node to a related document slice.

[0086] Step S140: Receive user query request and respond to user query request by performing in-depth analysis of user input data to obtain user query intent and query elements.

[0087] Specifically, the system receives user query requests and responds by performing in-depth analysis of the user input data to extract the user's query intent and query elements.

[0088] Step S150: Construct a condition graph based on the user's query intent and query elements, and map the condition graph to the knowledge network for entity localization and condition graph matching to generate structured search results.

[0089] Specifically, the system constructs a condition graph based on the user's query intent and query elements, and maps the constructed condition graph to the knowledge network for entity localization and condition graph matching to generate and output structured search results.

[0090] In some embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by the present invention includes the following steps in step S150:

[0091] Step S151: Construct a condition graph using the user's query intent and the query entities in the query elements as nodes, the entity relationships between different query entities as edges, and the entity attributes of the query entities as constraints.

[0092] Step S152: Calculate the vector similarity between entity nodes in the condition graph and entity nodes in the fused knowledge body, and locate the anchor entity between the condition graph and the fused knowledge body based on the vector similarity.

[0093] Step S153: Starting from the anchor entity, search for multiple knowledge subgraphs in the knowledge network that match the topological structure and semantic tags of the condition graph, and aggregate and semantically enhance the multiple knowledge subgraphs to obtain structured retrieval results.

[0094] Combination Figure 5 As shown in the specific embodiments, the multi-source information knowledge fusion and intelligent retrieval method provided by this invention relies on a unified intelligent retrieval technology based on conditional graphs for its knowledge fusion and inference engine. It utilizes the central capabilities of a large model to perform deep semantic analysis on user natural language queries to identify user query intent, entities, relationships, and constraints, dynamically generating a virtual conditional graph query structure. The nodes and edges of the conditional graph query structure represent entities and relational constraints in the query. Using the conditional graph as a whole query unit, graph matching and semantic search are performed directly on the fused knowledge body, achieving virtualized cross-database queries and replacing the multi-engine scheduling in existing technologies. Finally, the large model uses the matched knowledge subgraph and related document slices retrieved through bidirectional links to synthesize the final retrieval answer. The intelligent service and interaction layer provides natural language understanding interfaces, graph query and visualization engine services, and personalized recommendation and push engine services through a service bus or unified API gateway. The intelligent application layer supports intelligent applications such as natural language question answering, intelligent report generation, personalized knowledge portals, visual association analysis, and decision support systems.

[0095] During the user query and condition graph construction process, the user's natural language query is compiled into machine-executable, structured graph-pattern query instructions, including two processes: deep semantic parsing and dynamic condition graph generation. Deep semantic parsing utilizes a large language model to deeply understand the query, completing intent classification and element extraction. Intent classification determines whether the query belongs to a fact query, relational reasoning, list query, or comparative analysis type, providing guidance for subsequent construction. Element extraction accurately identifies the entities, relationships, attributes, and constraints mentioned in the query. Then, dynamic condition graph generation uses the extracted elements as a blueprint, and the system constructs a condition graph in memory in real time. Nodes in the condition graph represent entities or concepts in the query, with constraints such as name and type; edges represent expected relationships or logical connections between entities, with constraints such as type and direction. Attributes and constraints serve as additional conditions for nodes or edges. This condition graph is a virtual, temporary query structure, its form entirely determined by the current query, and is fundamentally different from the underlying static knowledge graph.

[0096] Next, the virtual query graph is mapped onto the actual knowledge network to perform a search, including two processes: semantic anchor point localization and graph pattern matching and reasoning. During semantic anchor point localization, entity nodes in the condition graph are quickly linked to their corresponding standardized entity nodes in the fused knowledge body using vector similarity calculation or a name dictionary to find the anchor point of the query in the knowledge network. During graph pattern matching and reasoning, starting from the anchor entity, an extended search of the graph database is performed. The entire condition graph is treated as a pattern, and subgraphs matching its topology and semantic labels are searched within the knowledge network. If the condition graph implies multi-hop relationships, a restricted graph traversal is performed in the knowledge network to find reasonable paths connecting the starting points. Simultaneously, during the matching process, attribute constraints are applied in real-time to filter candidate nodes and edges.

[0097] Finally, the structured results obtained from the matching are transformed into a user-understandable form, including three processes: candidate result aggregation, enhanced context acquisition, and structured result generation. Candidate result aggregation is used to aggregate multiple possible matching answer subgraphs. During enhanced context acquisition, the bidirectional links established during RAG and intelligent document slicing are used to automatically acquire original document slices related to entities in the answer subgraphs, providing detailed textual evidence for answer generation. During structured result generation, different forms of results are flexibly output according to the query intent. For simple facts (such as the attributes of an entity), values ​​are returned directly, generating a direct answer. For relational queries, a structured subgraph is returned, represented in a visual answer subgraph or a machine-readable JSON format. For complex analytical queries, the answer subgraph and related document slices are used as context input to the large language model, thereby generating a logically coherent and evidence-based natural language answer, i.e., a reasoning-based text report.

[0098] In this embodiment, a pre-constructed, intrinsically related fused knowledge body is built through unified knowledge modeling and multimodal semantic fusion before retrieval. During retrieval, the condition graph is directly matched with this unified knowledge network to obtain an answer subgraph with consistent structure and complete semantic association. Combined with intelligent document slicing and bidirectional links between knowledge nodes, the returned results possess both structured knowledge and original textual evidence, fundamentally solving the problem of result fusion and significantly improving the accuracy and information richness of the answers. Furthermore, through end-to-end direct mapping from natural language to the condition graph and then to the knowledge network, the technical complexity of all underlying data sources is shielded from users. Users only need to ask questions in natural language to directly access the deeply fused, full-domain knowledge and obtain more accurate search answers, revolutionizing the information retrieval service model. This allows even non-professional users to conduct complex knowledge exploration and association analysis, improving the usability and universality of the retrieval system.

[0099] The aforementioned multi-source information knowledge fusion and intelligent retrieval method acquires multi-source heterogeneous data and performs unified knowledge modeling on this data to obtain a unified knowledge framework. Then, based on this unified knowledge framework, a large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples. Next, attribute graphs are used to further structure these knowledge triples, and an embedding model is used to vectorize and store the associated data of each entity, resulting in a fused knowledge body. Subsequently, user query requests are received and responded to by performing deep analysis of the user input data to obtain the user's query intent and query elements. Finally, a condition graph is constructed based on the user's query intent and query elements, and this condition graph is mapped to the knowledge network for entity localization and condition graph matching to generate structured retrieval results. This method introduces a conditional graph as a unified virtual query interface, directly compiling the user's query intent into a graph structure query. This query can be directly performed on a unified fused knowledge body in the background for one-time graph pattern matching and semantic search. This effectively eliminates the network overhead and scheduling delay caused by multi-engine calls in existing technologies, simplifies the system architecture, and improves the retrieval response speed and the accuracy of retrieval results.

[0100] The multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is described below. The multi-source information knowledge fusion and intelligent retrieval system described below can be referred to in correspondence with the multi-source information knowledge fusion and intelligent retrieval method described above.

[0101] like Figure 6 As shown, in one embodiment, a multi-source information knowledge fusion and intelligent retrieval system includes a knowledge modeling module 610, a structured processing module 620, a knowledge fusion module 630, a query analysis module 640, and an intelligent retrieval module 650.

[0102] The knowledge modeling module 610 is used to acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework.

[0103] The structured processing module 620 is used to perform structured preprocessing on multi-source heterogeneous data based on a unified knowledge framework and by calling a large language model to generate structured knowledge triples.

[0104] The knowledge fusion module 630 is used to perform structuring processing on structured knowledge triples through attribute graphs and to vectorize and store the associated data of each entity through an embedding model to obtain a fused knowledge body.

[0105] The query analysis module 640 is used to receive user query requests and respond to them by performing in-depth analysis of the user input data to obtain the user's query intent and query elements.

[0106] The intelligent retrieval module 650 is used to construct a condition graph based on the user's query intent and query elements, and to map the condition graph to the knowledge network for entity localization and condition graph matching to generate structured retrieval results.

[0107] In this embodiment, the knowledge modeling module 610 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is specifically used for:

[0108] Knowledge domain analysis is performed on multi-source heterogeneous data to construct a hierarchical domain ontology based on the multi-source heterogeneous data. The domain ontology includes entity types, entity attributes, and entity relationships of the target knowledge domain.

[0109] The domain ontology is transformed into a knowledge graph containing entity types, entity attributes and entity relationships corresponding to each entity type, and multi-source heterogeneous data is standardized based on the knowledge graph to build a unified knowledge framework.

[0110] In this embodiment, the structured processing module 620 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is specifically used for:

[0111] Under a unified knowledge framework, a large language model is used to perform entity recognition, entity relation extraction, and entity attribute extraction on text data from multi-source heterogeneous data. The knowledge domain corresponding to the current entity is written into the model prompt words to generate structured knowledge triples.

[0112] The large language model transforms tabular data from multi-source heterogeneous data into tabular descriptive text, and then maps the tabular descriptive text to target entities and entity attributes in the knowledge domain.

[0113] In this embodiment, the structured processing module 620 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is further used for:

[0114] Text data is extracted from image data in multi-source heterogeneous datasets using OCR or parsing libraries, and the image description text is parsed. The text data and image description text in the image data are then input into a large language model to generate structured knowledge triples.

[0115] The large language model is used to perform entity implicit relation reasoning and knowledge completion processing on structured knowledge triples, and conflict detection is performed on structured knowledge triples to resolve the detected conflicts by calling preset strategies.

[0116] In this embodiment, the knowledge fusion module 630 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is specifically used for:

[0117] The original documents are extracted from multi-source heterogeneous data, and the semantic boundaries of the original documents are identified by natural language processing technology. Based on the semantic boundaries, the original documents are sliced ​​and metadata is annotated to obtain multiple document slices carrying metadata.

[0118] Extract the target entities from the document slices and establish a bidirectional index between the target entities and the fused knowledge body, so as to set semantic tags for the document slices based on the bidirectional index.

[0119] In this embodiment, the knowledge fusion module 630 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is further used for:

[0120] Each entity in the structured knowledge triple and the entity relationships between different entities are transformed into an attribute graph, and the attribute graph is stored in a graph database.

[0121] The embedding model transforms the entity description text, entity attributes, and document slices of each entity in the structured knowledge triple into high-dimensional vectors, and stores the high-dimensional vectors in a vector database.

[0122] In this embodiment, the intelligent retrieval module 650 of the multi-source information knowledge fusion and intelligent retrieval system provided by the present invention is specifically used for:

[0123] A conditional graph is constructed using the user's query intent and the query entities in the query elements as nodes, the entity relationships between different query entities as edges, and the entity attributes of the query entities as constraints.

[0124] Calculate the vector similarity between entity nodes in the conditional graph and entity nodes in the fused knowledge body, and locate the anchor entity between the conditional graph and the fused knowledge body based on the vector similarity.

[0125] Starting with the anchor entity, multiple knowledge subgraphs that match the topological structure and semantic tags of the condition graph are searched in the knowledge network. These subgraphs are then aggregated and semantically enhanced to obtain structured retrieval results.

[0126] Figure 7 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 7As shown. The electronic device includes a processor, internal memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-source information knowledge fusion and intelligent retrieval method, which includes:

[0127] Acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework;

[0128] Based on a unified knowledge framework, a large language model is invoked to perform structured preprocessing on multi-source heterogeneous data to generate structured knowledge triples.

[0129] The structured knowledge triples are structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body.

[0130] Receive user query requests and respond to them by performing in-depth analysis of user input data to obtain user query intent and query elements;

[0131] A condition graph is constructed based on the user's query intent and query elements, and then mapped to the knowledge network for entity localization and condition graph matching to generate structured search results.

[0132] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements a multi-source information knowledge fusion and intelligent retrieval method, the method comprising:

[0134] Acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework;

[0135] Based on a unified knowledge framework, a large language model is invoked to perform structured preprocessing on multi-source heterogeneous data to generate structured knowledge triples.

[0136] The structured knowledge triples are structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body.

[0137] Receive user query requests and respond to them by performing in-depth analysis of user input data to obtain user query intent and query elements;

[0138] A condition graph is constructed based on the user's query intent and query elements, and then mapped to the knowledge network for entity localization and condition graph matching to generate structured search results.

[0139] On another front, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and when the processor executes the computer instructions, it implements a multi-source information knowledge fusion and intelligent retrieval method, which includes:

[0140] Acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework;

[0141] Based on a unified knowledge framework, a large language model is invoked to perform structured preprocessing on multi-source heterogeneous data to generate structured knowledge triples.

[0142] The structured knowledge triples are structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body.

[0143] Receive user query requests and respond to them by performing in-depth analysis of user input data to obtain user query intent and query elements;

[0144] A condition graph is constructed based on the user's query intent and query elements, and then mapped to the knowledge network for entity localization and condition graph matching to generate structured search results.

[0145] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0146] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0147] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0148] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for multi-source information knowledge fusion and intelligent retrieval, characterized in that, The method includes: Acquire multi-source heterogeneous data and perform unified knowledge-based modeling on the multi-source heterogeneous data to obtain a unified knowledge framework; Based on the unified knowledge framework, the large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples. The structured knowledge triples are structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body. Receive user query requests and respond to the user query requests by performing in-depth analysis of user input data to obtain user query intent and query elements; A condition graph is constructed based on the user's query intent and query elements, and the condition graph is mapped to a knowledge network for entity localization and condition graph matching to generate structured search results; The structured knowledge triples are processed using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body, including: The original documents are extracted from the multi-source heterogeneous data, and the semantic boundaries of the original documents are identified by natural language processing technology. The original documents are then sliced ​​and metadata is annotated according to the semantic boundaries to obtain multiple document slices carrying metadata. Extract the target entity from the document slice and establish a bidirectional index between the target entity and the fused knowledge body, so as to set semantic tags for the document slice according to the bidirectional index; Each entity in the structured knowledge triple and the entity relationships between different entities are transformed into an attribute graph, and the attribute graph is stored in a graph database; The entity description text, entity attributes, and document slices of each entity in the structured knowledge triple are transformed into high-dimensional vectors by an embedding model, and the high-dimensional vectors are stored in a vector database. The step of constructing a condition graph based on the user's query intent and query elements, and mapping the condition graph to a knowledge network for entity localization and condition graph matching to generate structured search results includes: The conditional graph is constructed using the user's query intent and the query entities in the query elements as nodes, the entity relationships between different query entities as edges, and the entity attributes of the query entities as constraints. Calculate the vector similarity between entity nodes in the condition graph and entity nodes in the fused knowledge body, and locate the anchor entity between the condition graph and the fused knowledge body based on the vector similarity; Starting from the anchor entity, multiple knowledge subgraphs that match the topological structure and semantic tags of the condition graph are searched in the knowledge network, and the multiple knowledge subgraphs are aggregated and semantically enhanced to obtain the structured retrieval results.

2. The multi-source information knowledge fusion and intelligent retrieval method according to claim 1, characterized in that, The process of acquiring multi-source heterogeneous data and performing unified knowledge-based modeling on the multi-source heterogeneous data to obtain a unified knowledge framework includes: Knowledge domain analysis is performed on the multi-source heterogeneous data to construct a hierarchical domain ontology based on the multi-source heterogeneous data. The domain ontology includes entity types, entity attributes, and entity relationships of the target knowledge domain. The domain ontology is transformed into a knowledge graph containing entity types and corresponding entity attributes and entity relationships for each entity type. The multi-source heterogeneous data is then standardized based on the knowledge graph to construct the unified knowledge framework.

3. The multi-source information knowledge fusion and intelligent retrieval method according to claim 1, characterized in that, Based on the unified knowledge framework, the large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples, including: Under the unified knowledge framework, entity recognition, entity relation extraction and entity attribute extraction are performed on the text data in the multi-source heterogeneous data through the large language model, and the knowledge domain corresponding to the current entity is written into the model prompt words to generate the structured knowledge triples. The tabular data in the multi-source heterogeneous data is transformed into tabular description text through a large language model, and the tabular description text is mapped to the target entities and entity attributes in the knowledge domain.

4. The multi-source information knowledge fusion and intelligent retrieval method according to claim 3, characterized in that, The step of using the unified knowledge framework and calling a large language model to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples also includes: Text data is extracted from the image data in the multi-source heterogeneous data using OCR or a parsing library, and the image description text is parsed. The text data and image description text in the image data are then input into a large language model to generate the structured knowledge triples. The structured knowledge triples are subjected to entity implicit relation reasoning and knowledge completion processing using a large language model, and conflict detection is performed on the structured knowledge triples to resolve the detected conflicts by invoking a preset strategy.

5. A multi-source information knowledge fusion and intelligent retrieval system, characterized in that, The system is used to implement the multi-source information knowledge fusion and intelligent retrieval method according to any one of claims 1 to 4, the system comprising: The knowledge modeling module is used to acquire multi-source heterogeneous data and perform unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework. The structured processing module is used to perform structured preprocessing on the multi-source heterogeneous data by calling the large language model based on the unified knowledge framework, so as to generate structured knowledge triples. The knowledge fusion module is used to perform structuring processing on the structured knowledge triples through attribute graphs, and to vectorize and store the associated data of each entity through an embedding model to obtain a fused knowledge body. The query analysis module is used to receive user query requests and, in response to the user query requests, perform in-depth analysis of user input data to obtain user query intent and query elements; The intelligent retrieval module is used to construct a condition graph based on the user's query intent and query elements, and to map the condition graph to a knowledge network for entity localization and condition graph matching to generate structured retrieval results.

6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-source information knowledge fusion and intelligent retrieval method as described in any one of claims 1 to 4.

7. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-source information knowledge fusion and intelligent retrieval method as described in any one of claims 1 to 4.

Images