An enhanced retrieval method and system based on ranking of graph node importance

Abstract

The application discloses an enhanced retrieval method and system based on graph node importance ranking, wherein the method comprises the following steps: performing semantic abstract extraction on original documents, constructing entity-relation triplets, and generating abstract nodes and entity relation nodes; performing semantic vector coding on a user query statement, constructing a candidate document subgraph based on a preliminary retrieval result and a knowledge graph; performing importance scoring and ranking on the nodes of the candidate document subgraph, extracting abstract nodes according to the ranking result, and generating output text for answering. The application realizes refined expression and relationship modeling of document information, provides a semantic index basis for subsequent retrieval, makes retrieval no longer limited to keyword matching, but focuses on the semantic association between the query and the entity, ensures that the screening result closely revolves around the user's intention, compresses the input length while ensuring semantic integrity, and effectively improves the relevance and accuracy of the generated result of the large model.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and knowledge augmentation generation, and in particular to an augmented retrieval method and system based on graph node importance ranking, as well as storage media and electronic devices. Background Technology

[0002] With the widespread application of Natural Language Processing (NLP) and large-scale pre-trained language models (such as BERT and GPT), text generation and question answering systems based on semantic understanding have gradually become the core technology path for intelligent information services. Traditional generation methods mostly rely on the Retrieval-Augmented Generation (RAG) framework, which retrieves relevant information fragments from the document library as prompts to input into the large language model to generate answers. At the same time, knowledge representation methods such as entity extraction, knowledge graph construction, and node ranking are also gradually being integrated into the language model architecture to improve semantic understanding and factual expression capabilities in complex tasks. This trend has driven the transition from pure semantics to structural semantics, enabling information-enhanced generation systems to evolve towards higher accuracy and interpretability.

[0003] Despite significant progress in semantic retrieval and text generation technologies, key bottlenecks remain when dealing with input scenarios involving long documents, multiple entities, and high-density knowledge points. Traditional RAG frameworks primarily rely on dense matching in vector spaces, retrieving relevant text segments solely through semantic similarity, neglecting the structured semantic relationships between entities within a document. This often results in redundant, one-sided, or fragmented recall results. Existing summarization technologies largely depend on the model's compression capabilities, lacking mechanisms for discerning the hierarchy of importance between information segments. This is particularly problematic in scenarios involving the fusion of multiple documents or summarization fragments, making it difficult to accurately assess the contribution of text segments. While graph neural networks and other technologies have been used for node ranking in knowledge graphs, they are not yet deeply coupled with the generation capabilities of large language models. The lack of a linkage mechanism between information filtering and summary output makes it difficult to transform ranking results into high-quality input prompts. Therefore, existing technologies cannot achieve a generation enhancement path oriented towards multiple documents and structures with a node importance assessment mechanism, lacking structure-oriented information fusion capabilities and failing to effectively optimize model performance in knowledge-intensive question-answering tasks. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing RAG enhancement generation technology suffers from poor structural understanding, low information filtering accuracy, insufficient long text summary fusion capability, and the problem of how to achieve high-quality summary information organization before large language model prompts are generated.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: an enhanced retrieval method based on graph node importance ranking, comprising performing semantic summary extraction on the original document, constructing entity-relation triples and generating summary nodes and entity-relation nodes; performing semantic vector encoding on the user query statement, constructing candidate document subgraphs based on preliminary retrieval results and knowledge graph; performing importance scoring and ranking on the nodes of the candidate document subgraphs, extracting summary nodes based on the ranking results and generating output text for answering.

[0007] Optionally, in the above-described method embodiments of the present invention, the step of performing semantic summary extraction on the original document includes: processing the content of the original input document using a large language model, performing semantic recognition and compressed summarization on the entire document, extracting sentences or paragraphs with core meaning from the text, and generating summary text of a limited length, with each summary covering the main content of the document; when the document length exceeds a preset threshold, automatically segmenting it by paragraph or semantic unit, and generating an independent summary for each segment.

[0008] Optionally, in the above-described method embodiments of the present invention, the construction of entity-relation triples and the generation of summary nodes and entity-relation nodes includes: using a large language model and text analysis algorithms to perform natural language parsing on each document; extracting meaningful information units through entity recognition, including names, organizations, time information, monetary values, and percentages; determining the behavioral or state relationships between different entities through syntactic analysis and semantic relationship recognition technology, and summarizing them into structured triple information, with each set of data consisting of a starting entity, a relationship type, and a target entity; storing the triple data uniformly in a knowledge graph structure, with each triple serving as an edge in the graph, connecting the corresponding entity nodes, and establishing an index connection relationship in the graph with the original document through the summary node.

[0009] Optionally, in the above-described method embodiments of the present invention, the step of semantically encoding the user query statement includes: after receiving the user's natural language input, converting the query statement into a semantic vector form of uniform dimension through an embedded encoding model, with the vector expression method adapting to both dense and sparse retrieval tasks; after the conversion is completed, performing similarity retrieval on the user's semantic vector through a pre-loaded document semantic vector library, and using a vector matching mechanism to return a set of documents with the highest similarity to the query; during the pre-retrieval process, based on the metadata tags corresponding to each document, including document timestamps, source types, or content identifiers, performing metadata-based filtering and re-sorting operations on the document results obtained from the preliminary retrieval, and adjusting the priority of the documents in the candidate set.

[0010] Optionally, in the above-described method embodiments of the present invention, the construction of the candidate document subgraph includes, after obtaining a preliminarily matched document set, filtering out all entity nodes associated with the corresponding summary nodes in the document set and their connection relationships in the graph from the knowledge graph structure, constructing a candidate subgraph within a limited scope, the candidate subgraph including document summary nodes, entity nodes, and connection nodes composed of triple relationships, expressing the combination relationship of document content in semantic structure through structured links of nodes and edges; after the candidate subgraph is constructed, performing a node filtering operation based on the number of connections between nodes and the degree of semantic consistency with the summary nodes, reducing the number of redundant entity nodes, and retaining only some nodes that have multi-edge connection relationships with the summary nodes or have high semantic relevance.

[0011] Optionally, in the above-described method embodiments of the present invention, the step of performing importance scoring and ranking on candidate document subgraph nodes includes assigning an initial importance score value to each node in the candidate subgraph, and by traversing the connection relationships between nodes, weighting and propagating the importance score information from adjacent nodes according to the number of connections and link strength, and accumulating and updating it to the current node; the importance scoring is performed through multiple rounds of iterative propagation, and each round of iteration updates each node based on the importance score result of the previous round; when the change in the node's importance score meets the preset convergence condition or reaches the maximum number of iterations, the current importance score state is used as the final ranking basis.

[0012] Optionally, in the above-described method embodiments of the present invention, the step of extracting summary nodes and generating output text for answering questions based on the ranking results includes: after ranking the importance of nodes, selecting the top-scoring article nodes from the ranking results and extracting the corresponding summary node content as a candidate summary set; for article nodes originating from the original documents that have been split and processed, sequentially splicing multiple summary fragments of the current document according to the paragraph order in the original document, ensuring that the content structure after splicing remains consistent with the original text and avoiding paragraph information confusion; during the construction of the summary set, calculating the total length of all summary content and handling cases that exceed the preset input limit, when the total length of the summary exceeds the maximum number of input characters supported by the large language model, filtering or compressing the summary content according to the score priority of the article nodes in the ranking, confirming that the length of the final summary set is controlled and the semantic coverage is complete; the processed summary set will be submitted to the large language model as the final input to generate the answer content for the user's question.

[0013] According to another aspect of the present invention, an enhanced retrieval system based on graph node importance ranking is provided, including a semantic extraction module, a vector retrieval module, and a graph ranking module;

[0014] The semantic extraction module is used to perform semantic summary extraction on the original document, construct entity-relation triples and generate summary nodes and entity-relation nodes;

[0015] The vector retrieval module is used to encode the user's query statement into semantic vectors and construct a candidate document subgraph based on the preliminary retrieval results and the knowledge graph.

[0016] The graph sorting module is used to perform importance scoring and sorting on candidate document subgraph nodes, extract summary nodes based on the sorting results, and generate output text for answering questions.

[0017] The beneficial effects of this invention are as follows: The enhanced retrieval method based on graph node importance ranking provided by this invention performs semantic summary extraction on the original document, constructs entity-relation triples and generates summary nodes and entity-relation nodes, realizing the refined expression of document information and relation modeling, providing a semantic index foundation for subsequent retrieval. It encodes user query statements into semantic vectors, constructs candidate document subgraphs based on preliminary retrieval results and knowledge graphs, achieving high-dimensional matching between natural language queries and semantic documents. It performs importance scoring and ranking on the nodes of the candidate document subgraphs, extracts summary nodes based on the ranking results and generates output text for the answer, ensuring semantic integrity while compressing the input length, effectively improving the relevance and accuracy of the large model's generated results. Therefore, this invention achieves structured expression of document content, efficient matching of user queries and document semantics, and accurate extraction of high-quality summary information. It solves the shortcomings of existing technologies in terms of insufficient unstructured text processing capabilities, low semantic retrieval accuracy, and redundant and irrelevant generated input content. By guiding the selection of summary content through graph node importance ranking results, it achieves a precise correspondence between generated content and semantic structure, improving the contextual relevance and content representativeness of the output results.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same parts or steps.

[0020] Figure 1 This is a flowchart illustrating a method provided in an exemplary embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the system provided in an exemplary embodiment of the present invention.

[0022] Figure 3 This is a flowchart of LLM-based knowledge graph construction provided by an exemplary embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of knowledge graph reordering provided by an exemplary embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of an enhanced generation of retrieval results based on graph node importance ranking provided by an exemplary embodiment of the present invention. Detailed Implementation

[0025] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0026] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0027] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of the present invention are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0028] It should also be understood that in the embodiments of the present invention, "multiple" can refer to two or more, and "at least one" can refer to one, two or more.

[0029] It should also be understood that any component, data or structure mentioned in the embodiments of the present invention can generally be understood as one or more unless explicitly defined or given contrary instructions in the context.

[0030] refer to Figure 1 As an embodiment of the present invention, an enhanced retrieval method based on graph node importance ranking is provided, comprising:

[0031] Step 101: Perform semantic summary extraction on the original document, construct entity-relation triples and generate summary nodes and entity-relation nodes.

[0032] Furthermore, semantic summarization extraction of the original document includes processing the original input document content using a large language model, performing semantic recognition and compression summarization on the entire document, extracting sentences or paragraphs with core meaning from the text, and generating summary text of a limited length, with each summary covering the main content of the document; when the document length exceeds a preset threshold, it is automatically segmented by paragraph or semantic unit, and an independent summary is generated for each segment.

[0033] It should be noted that constructing entity-relation triples and generating summary nodes and entity-relation nodes involves: using large language models and text analysis algorithms to perform natural language parsing on each document; extracting meaningful information units through entity recognition, including names, organizations, time information, monetary values, and percentages; determining the behavioral or state relationships between different entities through syntactic analysis and semantic relation recognition techniques, and summarizing them into structured triple information, with each set of data consisting of a starting entity, relation type, and target entity; storing the triple data uniformly in a knowledge graph structure, with each triple serving as an edge in the graph, connecting the corresponding entity nodes, and establishing an index connection relationship in the graph with the original document through the summary node.

[0034] It should also be noted that by performing semantic summarization on the original document and constructing entity-relation triples, representative semantic fragments and structured knowledge can be extracted from large-scale text information. Summarization extracts lengthy documents into core sentences, allowing subsequent graph construction and querying to focus on key information and reducing processing costs. Triple construction, through entity recognition and relation extraction, transforms linguistic expressions into machine-readable structured data, forming a knowledge representation model for the document. This step not only enhances the system's information abstraction capabilities but also constructs a basic framework that supports semantic association and graph structure traversal, providing semantic structure support for downstream retrieval and ranking. It realizes the transformation from unstructured text to structured knowledge graphs, laying the foundation for subsequent accurate semantic positioning and graph-based retrieval.

[0035] Step 102: Encode the user query statement into a semantic vector and construct a candidate document subgraph based on the preliminary search results and the knowledge graph.

[0036] Furthermore, semantic vector encoding of user queries involves receiving the user's natural language input and converting the query into a unified-dimensional semantic vector form using an embedded encoding model. This vector representation is suitable for both dense and sparse retrieval tasks. After conversion, a pre-loaded document semantic vector library is used to perform similarity retrieval on the user's semantic vectors, and a vector matching mechanism is employed to return a set of documents with the highest similarity to the query. During the pre-retrieval process, based on the metadata tags corresponding to each document, including document timestamps, source types, or content identifiers, metadata-based filtering and re-sorting operations are performed on the initially retrieved document results to adjust the priority of documents in the candidate set.

[0037] It should be noted that constructing the candidate document subgraph involves, after obtaining the initially matched document set, filtering out all entity nodes associated with the corresponding summary nodes in the document set and their connection relationships in the graph from the knowledge graph structure, constructing a limited candidate subgraph. The candidate subgraph contains document summary nodes, entity nodes, and connection nodes composed of triple relationships. The structured links between nodes and edges express the combination relationship of document content in the semantic structure. After the candidate subgraph is constructed, a node filtering operation is performed based on the number of connections between nodes and the degree of semantic consistency with the summary nodes to reduce the number of redundant entity nodes, retaining only those nodes that have multiple edge connections with the summary nodes or have high semantic relevance.

[0038] It should also be noted that by semantically encoding user queries and constructing candidate document subgraphs from the knowledge graph based on preliminary retrieval results, precise localization from user intent to semantic context is achieved. Vector encoding endows queries with high-dimensional semantic representation capabilities, making them suitable for various retrieval strategies. Combining similarity matching with the document semantic vector library effectively filters out potentially relevant documents, and further adjusts the ranking through metadata to enhance contextual matching capabilities. Subsequently, candidate subgraphs are constructed based on the knowledge graph, retaining nodes and semantic paths highly relevant to user intent, thus realizing the construction of a query-oriented semantic field. This not only improves the relevance and accuracy of retrieval but also introduces structural semantic information, enhancing the semantic focus and interpretability of subsequent ranking and generation.

[0039] Step 103: Perform importance scoring and ranking on the candidate document subgraph nodes, extract summary nodes based on the ranking results, and generate output text for the answer.

[0040] Furthermore, the importance scoring and ranking of candidate document subgraph nodes includes assigning an initial importance score to each node in the candidate subgraph, and traversing the connection relationships between nodes to propagate the importance score information from adjacent nodes in a weighted manner according to the number of connections and link strength, and accumulating and updating it to the current node; the importance scoring is performed through multiple rounds of iterative propagation, with each round updating each node based on the importance score result of the previous round, and when the change in the node's importance score meets the preset convergence condition or reaches the maximum number of iterations, the current importance score state is used as the final ranking basis.

[0041] It should be noted that the process of extracting summary nodes and generating output text for answers based on the ranking results includes: after ranking the importance of nodes, selecting the top-scoring article nodes from the ranking results and extracting their corresponding summary node content as a candidate summary set; for article nodes originating from the split original documents, sequentially concatenating multiple summary fragments of the current document according to the paragraph order in the original document, ensuring that the structure of the concatenated content remains consistent with the original text and avoiding paragraph information confusion; during the construction of the summary set, the total length of all summary content is counted, and cases exceeding the preset input limit are handled. When the total length of the summary exceeds the maximum number of input characters supported by the large language model, the summary content is filtered or compressed according to the score priority of the article nodes in the ranking to ensure that the final summary set has a controlled length and complete semantic coverage; the processed summary set will be submitted to the large language model as the final input to generate the answer content for the user's question.

[0042] It should also be noted that by performing importance scoring and ranking on each node in the candidate subgraph, and extracting summary nodes based on the ranking results to generate output text, high-value information can be identified from the graph structure and highly relevant answer content can be constructed. The scoring mechanism is based on iterative propagation of the number of connections and link strength, so that the importance of a node not only reflects information density but also its influence in the semantic network. The ranked summary nodes serve as the input for generation, and their content is sequentially concatenated at the paragraph level with length control to ensure that the information structure is coherent and does not exceed the limit. The final summary set has the characteristics of semantic concentration and concise content, which improves the quality and stability of the large model's generated results, achieves accurate mapping from structured information to language generation, and enhances the user's question-answering experience and the system's intelligent expression capabilities.

[0043] Reference Figure 2 As an embodiment of the present invention, an enhanced retrieval system based on graph node importance ranking is provided, including a semantic extraction module 21, a vector retrieval module 22, and a graph ranking module 23.

[0044] The semantic extraction module 21 is used to extract semantic summaries from the original documents, construct entity-relation triples and generate summary nodes and entity-relation nodes; the vector retrieval module 22 is used to encode the user query statement into semantic vectors and construct candidate document subgraphs based on the preliminary retrieval results and knowledge graph; the graph ranking module 23 is used to perform importance scoring and ranking on the nodes of the candidate document subgraphs, extract summary nodes based on the ranking results and generate output text for answering questions.

[0045] The overall solution of this invention can be divided into three stages: knowledge graph construction, pre-retrieval, and knowledge graph reordering.

[0046] Knowledge graph construction stage, such as Figure 3 As shown, Figure 3 This is a flowchart of LLM-based knowledge graph construction provided by an exemplary embodiment of the present invention.

[0047] The purpose of knowledge graph construction is to leverage the capabilities of large language models to extract summaries, recognize entities, and extract triples from document-type information. This involves decomposing various original documents into summary information, entity recognition data, and triples, and storing these data in the knowledge graph. Specifically:

[0048] Abstract Information: This patent utilizes a large model to automatically condense the main content, core viewpoints, or key information of text into a shorter version. The abstract length is set at 400 characters, while striving to retain the essence and intent of the original text. Its core objective is to improve information retrieval efficiency and help users quickly understand the core content of lengthy texts. A [filename-abstract] triple is constructed and stored in the knowledge graph. This patent also employs special processing for extremely long texts. Due to the large amount of information in extremely long texts, this patent segments documents exceeding 3000 characters, such as dividing "Article" into "Article 1", "Article 2", ... "Article n", and performs separate abstract extraction and triple construction for each article. This better avoids information loss during the processing of extremely long texts.

[0049] Entity identification information: Automatically identify, locate, and classify named entities with specific meanings from unstructured text. The most common categories include: Person names (PER): such as "Zhang San," "Marie Curie"; Place names / locations (LOC): such as "Beijing," "Pacific Ocean"; Organization names (ORG): such as "Tsinghua University," "Microsoft Corporation"; Time expressions (TIME): such as "2023," "Next Monday," "221 BC"; Dates (DATE): such as "July 4th," "National Day"; Currency / amount (MONEY): such as "100 USD," "Fifty Million"; Percentages (PERCENT): such as "50%"; Others: Domain-specific entities such as "disease names" (diabetes), "drug names" (aspirin), "legal terms," ​​etc.

[0050] Information extraction from triples: Dependency parsing is performed using a large model to extract the relationships between two entities, such as [Newton - birth - 1643] and [Jiuquan - launch - Shenzhou V]. The extracted triple relationships are then stored in a knowledge graph.

[0051] The final result is a knowledge graph of multiple original documents, including document names, summaries, key document information, and their interrelationships. See the flowchart for the LLM-based knowledge graph construction process. Figure 3 .

[0052] Pre-retrieval stage: The pre-retrieval stage adopts the general BGE-m3 algorithm to convert the user's query into an embedding representation, realizing the efficient three-in-one function of dense retrieval, sparse retrieval and multi-vector retrieval. It achieves the pre-retrieval effect for the user's question. The number of documents returned by the pre-retrieval used in this patent is 500.

[0053] Knowledge graph reordering stage, such as Figure 4 As shown, Figure 4 This is a schematic diagram of knowledge graph reordering provided by an exemplary embodiment of the present invention.

[0054] First, the 500 documents retrieved in the pre-retrieval stage are used as initial weighted documents to construct a full-document subgraph. This involves selecting summaries, entities, and relationships related to the pre-retrieval result documents from the entire knowledge graph to construct a knowledge graph with a limited set.

[0055] Secondly, the PageRank algorithm is used to calculate the weight of each node in the knowledge graph, where d is the damping coefficient (0.85 in this patent), N is the total number of nodes in the knowledge graph subgraph, Ti is all nodes linked to node A, and C(Ti) represents the number of output edges of Ti.

[0056]

[0057] After calculating the PageRank (PR) values ​​for all nodes, the top 10 article nodes are selected (since the sub-knowledge graph contains both article and non-article nodes, but non-article nodes have no effect on the model's understanding output, only article nodes are selected), and the corresponding summary nodes of the article nodes are output as the final output, which is then submitted to the large language model. See the attached diagram for a detailed illustration of the knowledge graph reordering. Figure 2 .

[0058] Ultimately, the large model obtains accurate article summary information and combines this information to answer user questions. See [link to full workflow] for details. Figure 5 . Figure 5 This is a schematic diagram of an enhanced generation of retrieval results based on graph node importance ranking provided by an exemplary embodiment of the present invention.

[0059] The key technical points of this invention are: 1. It introduces both knowledge graph and vector retrieval approaches to construct an advanced RAG method, which improves the accuracy of large-scale model question-answering outputs for specific tasks (private domain document-based question answering) while avoiding the complex module combination process of modular RAG. 2. By adopting the PageRank approach, the traditional re-ranking process is transformed into a knowledge graph node importance ranking process, considering the correlation between documents and avoiding the singularity of traditional retrieval methods that only use sparse or dense retrieval methods to search only document content, greatly enhancing the accuracy of private domain knowledge retrieval in a specific professional field. 3. At the same time, since the final information submitted to the large model is the summary information of the segmented article, it avoids the excessive information length caused by full-text loading. Although some information is lost to some extent due to text summarization, the article content is compressed, greatly improving the richness of the content input to the large model, and the overall computing speed is also improved.

[0060] The retrieval enhancement generation technology based on graph node importance ranking provided by this invention has the following advantages: 1. This technology integrates the advantages of graph computing's association search with the speed of traditional vector retrieval. For retrieval queries of private domain knowledge documents such as policies, books, and regulations, it can simultaneously consider related information, resulting in high accuracy and speed. In actual experiments, the recall rate was improved by approximately 5%. 2. Simultaneously, by introducing the concept of related subgraphs and the PageRank algorithm, it can quickly filter and rank the pre-retrieval results to obtain related information relevant to the user's question, greatly avoiding the problems of large model illusion output and unclear dependencies. 3. By constructing a knowledge graph through segmentation and summarization, it effectively avoids information loss while highly condensing core content, facilitating large models to obtain more information, thereby better leveraging the understanding capabilities of large models and outputting more accurate correct answers.

[0061] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0062] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0063] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0064] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0065] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps are decomposable and / or recombinable. Such decomposition and / or recombination should be considered equivalent to the present disclosure. The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0066] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. An enhanced retrieval method based on graph node importance ranking, characterized in that, include: Perform semantic summary extraction on the original document, construct entity-relation triples and generate summary nodes and entity-relation nodes; The user query statement is semantically vectorized, and a candidate document subgraph is constructed based on the preliminary search results and the knowledge graph. The construction of the candidate document subgraph includes, after obtaining the preliminary matched document set, filtering out all entity nodes associated with the corresponding summary nodes in the document set and their connection relationships in the graph from the knowledge graph structure, and constructing a candidate subgraph within a limited scope. The candidate subgraph contains document summary nodes, entity nodes, and connection nodes composed of triple relationships. The structured links between nodes and edges express the combination relationship of document content in the semantic structure. After the candidate subgraph is constructed, a node filtering operation is performed based on the number of connections between nodes and the degree of semantic consistency with the summary node to reduce the number of redundant entity nodes and retain only some nodes that have multilateral connections with the summary node or have high semantic relevance. The importance score and ranking of candidate document subgraph nodes are performed. Based on the ranking results, summary nodes are extracted and output text for answering is generated. The step of extracting summary nodes and generating output text for answering based on the ranking results includes: after completing the importance ranking of nodes, selecting the top-scoring article nodes from the ranking results and extracting the corresponding summary node content as a candidate summary set; for article nodes originating from the original documents that have been split and processed, multiple summary fragments of the current document are sequentially spliced ​​together according to the paragraph order in the original document, and the content structure of the spliced ​​content remains consistent with the original text to avoid paragraph information disorder. During the construction of the summary set, the total length of all summary content is counted, and cases exceeding the preset input limit are handled. When the total length of the summary exceeds the maximum number of input characters supported by the large language model, the summary content is filtered or compressed according to the score priority of the article node in the sorting, to ensure that the length of the final summary set is controlled and the semantic coverage is complete. The processed summary set will be submitted to the large language model as the final input to generate the answer content of the user's question.

2. The enhanced retrieval method based on graph node importance ranking as described in claim 1, characterized in that: The semantic summary extraction of the original document includes processing the content of the original input document using a large language model, performing semantic recognition and compressed summarization on the entire document, extracting sentences or paragraphs with core meaning from the text, and generating a summary text of a limited length, with each summary covering the main content of the document; When the document length exceeds a preset threshold, it is automatically segmented by paragraph or semantic unit, and an independent summary is generated for each segment.

3. The enhanced retrieval method based on graph node importance ranking as described in claim 2, characterized in that: The construction of entity-relation triples and the generation of summary nodes and entity-relation nodes include: using a large language model and text analysis algorithm to perform natural language parsing on each document; extracting information units with practical significance through entity recognition; and recognizing objects including names, organizations, time information, monetary values, and percentages. By using syntax analysis and semantic relation recognition technology, the behavioral or state relationships between different entities are determined and summarized into structured triple information. Each set of data consists of the starting entity, relation type, and target entity. The triplet data is uniformly stored in the knowledge graph structure. Each triplet serves as an edge in the graph, connecting the corresponding entity nodes, and establishing an index connection relationship with the original document through the summary node.

4. The enhanced retrieval method based on graph node importance ranking as described in claim 3, characterized in that: The step of semantic vector encoding of user query statements includes receiving the user's natural language input and converting the query statement into a semantic vector form of a unified dimension through an embedded encoding model. The vector expression method is suitable for both dense and sparse retrieval tasks. After the conversion is completed, the similarity of the user's semantic vector is retrieved through the preloaded document semantic vector library, and the vector matching mechanism is used to return the set of documents with the highest similarity to the query. During the pre-retrieval process, based on the metadata tags corresponding to each document, including document timestamp, source type or content identifier, the document results obtained from the initial retrieval are filtered and re-sorted according to metadata, and the priority of the documents in the candidate set is adjusted.

5. The enhanced retrieval method based on graph node importance ranking as described in claim 1, characterized in that: The step of performing importance scoring and sorting on candidate document subgraph nodes includes assigning an initial importance score value to each node in the candidate subgraph, and by traversing the connection relationships between nodes, weighting and propagating the importance score information from adjacent nodes according to the number of connections and link strength, and accumulating and updating it to the current node. Importance scoring is performed through multiple rounds of iterative propagation. Each round of iteration updates each node based on the importance score results of the previous round. When the change in the importance score of a node meets the preset convergence condition or reaches the maximum number of iterations, the current importance score status is used as the final ranking basis.

6. An enhanced retrieval system based on graph node importance ranking, comprising a semantic extraction module, a vector retrieval module, and a graph ranking module; The semantic extraction module is used to perform semantic summary extraction on the original document, construct entity-relation triples and generate summary nodes and entity-relation nodes; The vector retrieval module is used to encode the user query statement into semantic vectors and construct a candidate document subgraph based on the preliminary retrieval results and the knowledge graph. The construction of the candidate document subgraph includes, after obtaining the preliminary matched document set, filtering out all entity nodes associated with the corresponding summary nodes in the document set and their connection relationships in the graph from the knowledge graph structure, and constructing a candidate subgraph within a limited scope. The candidate subgraph contains document summary nodes, entity nodes, and connection nodes composed of triple relationships. The structured links between nodes and edges express the combination relationship of document content in the semantic structure. After the candidate subgraph is constructed, a node filtering operation is performed based on the number of connections between nodes and the degree of semantic consistency with the summary node to reduce the number of redundant entity nodes and retain only some nodes that have multilateral connections with the summary node or have high semantic relevance. The graph sorting module is used to perform importance scoring and sorting on candidate document subgraph nodes, extract summary nodes based on the sorting results, and generate output text for answering questions. The step of extracting summary nodes based on the sorting results and generating output text for answering questions includes: after completing the importance sorting of nodes, selecting the top-scoring article nodes from the sorting results and extracting the corresponding summary node content as a candidate summary set; for article nodes originating from the original documents that have been split and processed, sequentially splicing multiple summary fragments of the current document according to the paragraph order in the original document, ensuring that the content structure of the spliced ​​content remains consistent with the original text and avoiding paragraph information disorder. During the construction of the summary set, the total length of all summary content is counted, and cases exceeding the preset input limit are handled. When the total length of the summary exceeds the maximum number of input characters supported by the large language model, the summary content is filtered or compressed according to the score priority of the article node in the sorting, to ensure that the length of the final summary set is controlled and the semantic coverage is complete. The processed summary set will be submitted to the large language model as the final input to generate the answer content of the user's question.

7. An electronic device, characterized in that, The electronic device includes: a memory and a processor, the memory and the processor being coupled; the memory stores program instructions that, when executed by the processor, cause the electronic device to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Includes a computer program that, when run on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 5.

Images