Knowledge graph construction method and system

Abstract

The application discloses a knowledge graph construction method and system, and relates to the technical field of knowledge graph construction.The knowledge graph construction method comprises the following steps: performing a streaming decoupling operation on a received unstructured document to obtain a pure text data stream and a visual data stream; performing semantic direct reading and extraction on the pure text data stream and the visual data stream to obtain a triple pool; performing induction and standardization processing on the triple pool to obtain standardized triple data; mapping the standardized triple data into nodes and directed edges; and constructing a knowledge graph according to the nodes and the directed edges.The knowledge graph construction method can improve the completeness and numerical accuracy of data extraction from a complex visual chart.

Description

Technical Field

[0001] This application relates to the field of knowledge graph construction technology, and more specifically, to a method and system for constructing knowledge graphs from unstructured documents. Background Technology

[0002] As large language models evolve towards multimodal approaches, constructing high-quality, large-scale training datasets has become a crucial prerequisite for improving model intelligence. In large-scale data preparation projects, data cleaning has long surpassed simple deduplication and format verification; its core objective has evolved into deep semantic extraction and structure transformation of massive amounts of unstructured heterogeneous data (such as PDF documents like financial research reports, industrial manuals, and technical standards).

[0003] Knowledge graphs, with their ability to accurately represent entity relationships, have become the standard technical paradigm for achieving "deep cleaning" and "structuring" of multimodal data. By transforming mixed unstructured documents into logically rigorous knowledge graphs, data preparation systems aim to remove redundant background noise and retain core business logic, thereby providing a high signal-to-noise ratio knowledge foundation for downstream question-answering systems and inference models.

[0004] However, existing multimodal data cleaning and graph construction techniques suffer from data fidelity issues when processing documents containing high-density charts and complex visual information, making it difficult for the cleaned data to support high-precision question-answering tasks.

[0005] The current mainstream cascaded processing scheme of "OCR to text - text description generation - entity extraction" is essentially a "lossy cleaning" strategy. This scheme forcibly compresses two-dimensional visual topological information (such as the slope changes of line charts, the area ratio of pie charts, and the path directions of flowcharts) into one-dimensional natural language text descriptions. This modality conversion process is accompanied by a huge loss of "information entropy".

[0006] For example, when describing complex "multi-dimensional year-on-year / month-on-month growth trends," natural language often suffers from problems such as vague expression, loss of precision, and even semantic ambiguity. When data cleaning relies on these vague intermediate texts for knowledge extraction, the generated structured data has actually lost the most critical quantitative features and logical chains of the original charts. This "semantic wear and tear during the cleaning process" directly leads to serious illusions and logical errors when downstream question-answering systems use this data to answer complex questions involving numerical comparisons and trend analysis, significantly reducing the model's inference accuracy.

[0007] The content of the background section is merely technology known to the public and does not necessarily represent existing technology in the field. Summary of the Invention

[0008] This application aims to provide a knowledge graph construction method and system to solve the data fidelity problem caused by semantic wear and tear in the existing data cleaning process.

[0009] According to one aspect of this application, a knowledge graph construction method is provided, comprising: performing a streaming decoupling operation on received unstructured documents to obtain a plain text data stream and a visual data stream; performing semantic direct reading and extraction on the plain text data stream and the visual data stream to obtain a triple pool; performing induction and standardization processing on the triple pool to obtain standardized triple data; mapping the standardized triple data to nodes and directed edges; and constructing a knowledge graph based on the nodes and directed edges.

[0010] According to some embodiments of this application, semantic direct reading and extraction of plain text data streams and visual data streams to obtain a triple pool includes: performing global perception description on the visual data stream to generate corresponding natural language description information; and performing mixed extraction of plain text data streams, visual data streams, and natural language description information to generate a triple pool.

[0011] According to some embodiments of this application, the process of summarizing and standardizing the triple pool to obtain standardized triple data includes: determining whether the original relation predicate of the triple pool is a known relation predicate; if so, mapping the original relation predicate to a standard relation predicate to obtain standardized triple data based on the standard relation predicate; if not, determining whether the original relation predicate meets a preset reliability criterion; if the original relation predicate meets the preset reliability criterion, determining the original relation predicate as a new relation predicate to obtain standardized triple data based on the new relation predicate.

[0012] According to some embodiments of this application, determining whether the original relational predicate meets the preset reliability criteria includes: determining whether the original relational predicate meets the preset reliability criteria based on the context information of the original relational predicate in the unstructured document.

[0013] According to some embodiments of this application, mapping standardized triple data to nodes and directed edges includes: aligning the standardized triple data to entities to obtain aligned standardized triple data; mapping the subjects and objects of the aligned standardized triple data to nodes; and mapping the relation predicates of the aligned standardized triple data to directed edges.

[0014] According to one aspect of this application, a knowledge graph construction system is provided. The knowledge graph construction system may include a document parsing engine module, a semantic direct reading operator module, an adaptive induction module, and a graph construction module. The document parsing engine module performs streaming decoupling on received unstructured documents to obtain a plain text data stream and a visual data stream; the semantic direct reading operator module performs semantic direct reading and extraction on the plain text data stream and the visual data stream to obtain a triple pool; the adaptive induction module performs induction and standardization processing on the triple pool to obtain standardized triple data; and the graph construction module maps the standardized triple data into nodes and directed edges, and constructs a knowledge graph based on the nodes and directed edges.

[0015] According to some embodiments of this application, the document parsing engine performs global perception description of the visual data stream and generates corresponding natural language description information; the document parsing engine performs mixed extraction of plain text data stream, visual data stream and natural language description information to generate a triple pool.

[0016] According to some embodiments of this application, the semantic direct reading operator module determines whether the original relation predicate of the triple pool is a known relation predicate; if so, the semantic direct reading operator module maps the original relation predicate to a standard relation predicate to obtain standardized triple data based on the standard relation predicate; if not, the semantic direct reading operator module determines whether the original relation predicate meets the preset reliability standard; if so, if the original relation predicate meets the preset reliability standard, the semantic direct reading operator module determines the original relation predicate as a new relation predicate to obtain standardized triple data based on the new relation predicate.

[0017] According to some embodiments of this application, the semantic direct reading operator module determines whether the original relational predicate meets the preset reliability criteria based on the context information of the original relational predicate in the unstructured document.

[0018] According to some embodiments of this application, the graph construction module performs entity alignment on the standardized triple data to obtain aligned standardized triple data; the graph construction module maps the subjects and objects of the aligned standardized triple data to nodes; and the graph construction module maps the relation predicates of the aligned standardized triple data to directed edges.

[0019] The technical solution of this application can obtain plain text data streams and visual data streams by performing streaming decoupling operations on unstructured documents. The technical solution of this application can obtain a triple pool by performing semantic direct reading and extraction on the plain text data streams and visual data streams. The technical solution of this application can obtain standardized triple data by summarizing and standardizing the triple pool. The technical solution of this application can map the standardized triple data into nodes and directed edges. The technical solution of this application can construct a knowledge graph using nodes and directed edges.

[0020] The technical solution of this application performs semantic reading and extraction of visual data stream directly in pixel space through semantic direct reading operators. By eliminating the intermediate text conversion step, it can completely preserve the trend line slope, color legend correspondence and spatial layout features in the chart, thereby significantly improving the integrity and numerical accuracy of data extracted from complex visual charts. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a knowledge graph construction method 1000 according to an embodiment of this application is shown. Figure 2 A flowchart illustrating step S200 according to an embodiment of this application is shown; Figure 3 A flowchart illustrating step S300 according to an embodiment of this application is shown; Figure 4 A flowchart illustrating step S400 according to an embodiment of this application is shown; Figure 5 This diagram illustrates the structure of a knowledge graph construction system according to an embodiment of this application. Explanation of reference numerals in the attached figures: 20. Knowledge graph construction system; 21. Document parsing engine module; 22. Semantic direct reading operator module; 23. Adaptive induction module; 24. Graph construction module. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that this application will be thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0024] The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of these specific details, or other methods, components, materials, devices, etc. In these cases, well-known structures, methods, devices, implementations, materials, or operations will not be shown or described in detail.

[0025] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0026] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order.

[0027] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0028] The English terms used in this application, their full English names, and their corresponding Chinese definitions are as follows: OCR, Optical Character Recognition; PDF, Portable Document Format; JSON, JavaScript Object Notation, is a JavaScript object representation. XML, or eXtensible Markup Language; id, Identifier; bbox, bounding box.

[0029] See Figure 5The knowledge graph construction system 20 provided in this application may include a document parsing engine module 21, a semantic direct reading operator module 22, an adaptive induction module 23, and a graph construction module 24. The knowledge graph construction method 1000 provided in this application is described below with reference to the figures. See also... Figure 1 The knowledge graph construction method 1000 includes steps S100-S500.

[0030] In step S100, the received unstructured document is subjected to a streaming decoupling operation to obtain a plain text data stream and a visual data stream.

[0031] According to the example embodiments, unstructured documents can be information carriers whose content is stored in a non-predefined format and does not follow a fixed data model. For example, unstructured documents include binary documents such as PDF research reports, industry manuals, and technical standards. Unstructured documents also include financial research report PDFs containing line charts, pie charts, flowcharts, and data tables, equipment manual PDFs containing technical diagrams, and academic papers with multi-column mixed text and graphics.

[0032] Streaming decoupling is a processing method that separates unstructured documents in real time according to information modality categories and outputs them to independent data channels, achieving physical isolation and independent transmission of information of different modalities.

[0033] Plain text data streams can be sequences of text content extracted from unstructured documents, excluding images and complex formatting, and continuously output using a standardized data structure. Formats for plain text data streams can include JSON sequences, XML node lists, and plain text paragraph streams. Each text data object in a plain text data stream can contain a unique identifier field (id), a text content field (text), a page number field (page), and a location coordinate field (bbox). The path field (path) is left empty to indicate that the node is of text type.

[0034] Visual data streams can be sequences of visual content extracted from unstructured documents and existing as image files. Each visual object includes its location context anchor information within the unstructured document. Visual data streams can include sequences of JSON placeholder objects centered around image file paths, inline sequences centered around hexadecimal encoded image data, etc. Each visual data object in a visual data stream contains a unique identifier field (id), a type marker field (image), an image physical storage path field (path), and an optional in-document location description field. The text field (text) can be empty or store auxiliary text information such as chart titles.

[0035] For example, in step S100, the document parsing engine module 21 can accurately identify different modal elements in the unstructured document based on the layout analysis algorithm, perform streaming decoupling operation, physically split the original unstructured document into a plain text data stream and a visual data stream (including charts and key images), and retain the original context anchor information for the visual objects.

[0036] The document parsing engine module 21 can be configured with a document parsing engine, which can be configured on a computing server with a high-performance graphics processor. The document parsing engine module 21 does not concern itself with the underlying OCR character recognition details; instead, it directly obtains the plain text data stream and the visual data stream by calling a high-precision document parsing engine interface.

[0037] For example, when the document parsing engine module 21 receives a PDF document to be processed, it performs a streaming decoupling operation on the PDF document. Based on the natural reading order of the PDF document, it breaks down the content of each page into a series of independent node objects and outputs them in the form of a JSON list. This list clearly preserves the logical order of the mixed text and images, while simultaneously achieving modal separation at the physical level.

[0038] For text content in a PDF document, the document parsing engine encapsulates it into text data objects with the type 'text'. Taking the document's first page as an example, the parsed JSON fragment contains title nodes and body paragraph nodes. The 'id' field of the text data object identifies its unique position in the sequence, the 'text' field carries the cleaned plain text content, and the 'path' field is left empty, thus obtaining a plain text data stream.

[0039] For visual content such as charts and photos in PDF documents, the document parsing engine performs a crucial image stripping operation. Instead of attempting to convert images into text, the engine directly extracts individual image files from the original page and stores them in a specified path on the server. In the output JSON sequence, the document parsing engine generates a visual data object tagged with type "image". The text field of the visual data object may be empty, but the path field fully records the physical storage address of the image file.

[0040] In step S200, semantic direct reading and extraction are performed on the plain text data stream and the visual data stream to obtain a triple pool.

[0041] According to the example implementation, a triple is the basic building block of a knowledge graph, consisting of three elements: Subject, Predicate, and Object. It expresses a structured semantic proposition that there is a certain relationship between one entity and another entity, and is usually represented in the form of (Subject, Predicate, Object).

[0042] A triple pool refers to a collection of raw triples that have not yet undergone normalization, dynamically accumulated and stored during the knowledge extraction process, serving as an intermediate cache for subsequent normalization processing. The triple pool can exist in the form of an in-memory list data structure, a message queue, or a temporary database table.

[0043] For example, in step S200, the semantic direct reading operator module 22 performs semantic direct reading and extraction on the plain text data stream and the visual data stream to obtain a triple pool. The semantic direct reading operator module 22 can be configured with semantic direct reading operators. The semantic direct reading operators can perform semantic understanding and knowledge extraction directly on the visual data object modality of the visual data stream without intermediate format conversion (such as OCR text recognition followed by text analysis).

[0044] The semantic direct reading operator module 22 performs semantic direct reading and extraction on the plain text data stream and the visual data stream, obtaining triples containing the subject, predicate, and object. After obtaining all the triples, the semantic direct reading operator module 22 forms a set, thus obtaining the triple pool.

[0045] In step S300, the triple pool is summarized and standardized to obtain standardized triple data.

[0046] According to the example implementation, standardized triplet data can be triplet data that has undergone induction and standardization processing.

[0047] For example, in step S300, the adaptive induction module 23 performs induction and standardization processing on the triple pool to obtain standardized triple data. The adaptive induction module 23 can operate independently of a predefined static ontology library, possessing semantic discrimination and induction capabilities, and can monitor the original relation predicates flowing into the triple pool in real time.

[0048] If the adaptive induction module 23 detects that the original relational predicates can be covered by the existing dynamic ontology library, the original relational predicates can be transformed into standard relational predicates to obtain standardized triplet data.

[0049] When the adaptive induction module 23 detects primitive relation predicates that cannot be covered by the existing dynamic ontology, it can evaluate their business value by combining the contextual information of the primitive relation predicates in the unstructured document. If the adaptive induction module 23 determines that the primitive relation predicate is valid business logic, it automatically generates standardized new relation definitions and updates the dynamic ontology. The adaptive induction module 23 uses the new relation definitions to convert the triple pool into standardized triple data, realizing adaptive expansion in the data structuring process.

[0050] In step S400, the standardized triplet data is mapped to nodes and directed edges.

[0051] According to the example embodiment, nodes can represent entities (including subjects and objects). Directed edges can represent relational predicates. For example, in step S400, the graph construction module 24 can map the subjects and objects of the normalized triplet data to nodes, and map the relational predicates of the normalized triplet data to directed edges.

[0052] In step S500, a knowledge graph is constructed based on the nodes and directed edges.

[0053] According to the example implementation, a knowledge graph is a semantic knowledge representation system that organizes entities and their relationships in a graph structure, consisting of a directed attribute graph composed of nodes and directed edges.

[0054] For example, in step S500, the graph construction module 24 can construct a knowledge graph based on nodes and directed edges. The graph construction module 24 is configured with a write interface adapted to a multimodal knowledge graph database, writing the knowledge graph to the multimodal knowledge graph database, completing the final persistence from unstructured documents to structured knowledge assets. The data from the unstructured documents is completely transformed into a queryable, reasonable, and traceable structured knowledge graph.

[0055] Through the above embodiments, the technical solution of this application can obtain a plain text data stream and a visual data stream by performing streaming decoupling operations on unstructured documents. The technical solution of this application can obtain a triple pool by performing semantic direct reading and extraction on the plain text data stream and the visual data stream. The technical solution of this application can obtain standardized triple data by summarizing and standardizing the triple pool. The technical solution of this application can map the standardized triple data into nodes and directed edges. The technical solution of this application can construct a knowledge graph using nodes and directed edges.

[0056] The technical solution of this application performs semantic reading and extraction of visual data stream directly in pixel space through semantic direct reading operators. By eliminating the intermediate text conversion step, it can completely preserve the trend line slope, color legend correspondence and spatial layout features in the chart, thereby significantly improving the integrity and numerical accuracy of data extracted from complex visual charts.

[0057] The technical solution of this application can overcome the bottleneck that the traditional cascaded solution of "OCR to text - text description generation - entity extraction" is prone to causing the loss of spatial topology and visual logic in the chart, and achieves high-fidelity retention of visual information.

[0058] Optionally, see Figure 2 Step S200 may include steps S210-S220.

[0059] In step S210, a global perception description is performed on the visual data stream to generate corresponding natural language description information.

[0060] According to the example implementation, global perception description can establish a structured natural language representation for visual data streams, thereby generating corresponding natural language description information.

[0061] For example, in step S210, the semantic direct reading operator module 22 performs global perception description on the visual data stream and generates corresponding natural language description information.

[0062] The semantic direct reading operator module 22 can send global perception commands to the multimodal large model, requiring it to act as a senior data scientist, perform pixel-level panoramic scanning of the visual data stream images, identify the type and core physical meaning of the chart, analyze the spatial dimensions represented by the coordinate axes, identify all visual entities in the image and their layout relationships, and finally summarize the scientific conclusions that the chart attempts to convey. After receiving the image and global perception commands, the multimodal large model outputs a natural language description containing rich semantics.

[0063] By performing a global perception description of the visual data stream, the spatial layout and geometric constraints of entities can be successfully captured, providing key semantic anchors for the next step of extraction.

[0064] In step S220, the plain text data stream, visual data stream, and natural language description information are mixed and extracted to generate a triple pool.

[0065] For example, in step S220, the semantic direct reading operator module 22 performs mixed extraction of plain text data stream, visual data stream and natural language description information to generate a triple pool.

[0066] The semantic direct reading operator module 22 can directly perform mixed extraction on plain text data streams to generate triples.

[0067] The semantic direct reading operator module 22 can extract a mixture of visual data stream and natural language description information, using both as input to perform fine-grained triple inference. During inference, it can focus on spatial topological relationships, compositional relationships, and logical inference relationships. The semantic direct reading operator module 22 quickly locates the pixel regions of each entity using locative words in the text description and verifies their relative positions by combining image features (such as image type, elements, style, trend, arrangement, key points, semantics, etc.), outputting a high-confidence list of triples that contains precise spatial relationships and abstract logical relationships.

[0068] With this configuration, natural language description information provides macro-level logical guidance, while visual data streams provide micro-level pixel evidence. The combination of the two generates a high signal-to-noise ratio triple pool, effectively solving the problems of easily lost details and logical breaks in complex charts.

[0069] Through the above embodiments, the technical solution of this application generates corresponding natural language description information by performing global perception description on the visual data stream. The technical solution of this application generates a triple pool by mixing and extracting plain text data stream, visual data stream and natural language description information.

[0070] The technical solution of this application adopts a "two-stage coarse-to-fine" perception, which significantly reduces the illusion rate of logical transformation. Through a cascade mechanism of global perception and hybrid extraction, the technical solution of this application uses natural language description information as a "thought chain" navigation to guide the extraction of triples to focus on the correct pixel area, effectively combining macro-logical guidance with micro-pixel evidence, and significantly improving the accuracy and robustness of complex logical relationship extraction.

[0071] Optionally, see Figure 3 Step S300 may include steps S310-S340.

[0072] In step S310, it is determined whether the original relation predicate of the triple pool is a known relation predicate.

[0073] According to the example embodiment, the known relational predicate can be a relational predicate stored in an existing dynamic ontology library. For example, in step S310, the adaptive induction module 23 performs a predicate matching operation to determine whether the original relational predicate in the triple pool is a known relational predicate.

[0074] If it is determined that the original relation predicate of the triple pool is a known relation predicate, step S320 can be executed.

[0075] In step S320, the original relational predicate is mapped to a standard relational predicate to obtain standardized triplet data based on the standard relational predicate.

[0076] According to the example embodiment, in step S320, the adaptive induction module 23 maps the original relation predicate to the standard relation predicate to obtain standardized triplet data based on the standard relation predicate.

[0077] For example, in the statistical chart of "Supermarket Composition in the Third Quarter" in the PDF document, the extracted triples include entities such as "domestic brands," "snack section," and "main shelves." The triples include primitive relational predicates such as "belongs to," "contains," and "occupies main shelves."

[0078] The adaptive induction module 23 determines that "belongs to" and "contains" belong to known relation predicates and directly maps them to standard relation predicates.

[0079] If it is determined that the original relation predicate of the triple pool is not a known relation predicate, step S330 can be executed.

[0080] In step S330, it is determined whether the original relational predicate meets the preset reliability criteria.

[0081] According to the example implementation, the pre-set reliability criteria can be used as a quantitative evaluation standard for the business value, semantic clarity and domain applicability of the original relational predicate.

[0082] In step S330, the adaptive induction module 23 determines whether the original relational predicate meets the preset reliability criteria. The adaptive induction module 23 can call the semantic analysis subroutine, and combine the context information of the original unstructured document containing the original relational predicate to comprehensively calculate the confidence score from three dimensions: business domain clarity, semantic unambiguity, and domain reusability value, to determine whether the original relational predicate meets the preset reliability criteria and make a registration or rejection decision.

[0083] Optionally, step S330 can specifically be: determining whether the original relational predicate meets the preset reliability criteria based on the context information of the original relational predicate in the unstructured document.

[0084] According to the example embodiment, the adaptive induction module 23 determines whether the original relational predicate meets the preset reliability criteria based on the context information of the original relational predicate in the unstructured document.

[0085] The adaptive induction module 23 constructs evaluation prompt words based on the original relational predicates. The confidence score of the adaptive induction module 23 and the document paragraphs in which it appears (with several sentences before and after as context information windows) is calculated comprehensively from three dimensions: business domain clarity, semantic unambiguity, and domain reusability value.

[0086] Business domain clarity assesses whether the original relational predicate describes a genuine business logic relationship, rather than an accidental linguistic expression; higher value results in a higher score. Semantic clarity assesses whether the original relational predicate has a single, clear semantic meaning in the current context; lower ambiguity results in a higher score. Domain reuse assesses whether the original relational predicate is likely to reappear in other documents within its business domain; higher reuse potential results in a higher score.

[0087] In step S340, if the original relation predicate meets the preset reliability criteria, the original relation predicate is determined as the new relation predicate, so as to obtain standardized triplet data based on the new relation predicate.

[0088] According to the example embodiment, if the original relational predicate meets the preset reliability criteria, the adaptive induction module 23 registers the original relational predicate as a new relational predicate so as to obtain standardized triplet data based on the new relational predicate.

[0089] For example, if the adaptive induction module 23 determines that "occupying the main shelf space" is not a known relation predicate, then the adaptive induction module 23 can evaluate it based on the contextual information of the chart corresponding to "occupying the main shelf space" in the PDF (i.e., supermarket merchandise display analysis). The adaptive induction module 23 analyzes that "occupying the main shelf space" describes a dominant spatial allocation state here, and has clear semantics and reusability value in the business domain of "retail / store management," meeting the pre-set reliability criteria. The adaptive induction module 23 registers "occupying the main shelf space" as a new relation predicate, thereby obtaining standardized triplet data.

[0090] Through the above embodiments, the technical solution of this application determines whether the original relation predicate of the triple pool is a known relation predicate. If it is determined that the original relation predicate of the triple pool is a known relation predicate, the original relation predicate is mapped to a standard relation predicate to obtain standardized triple data based on the standard relation predicate. If it is determined that the original relation predicate of the triple pool is not a known relation predicate, it is determined whether the original relation predicate meets a preset reliability criterion. If the original relation predicate meets the preset reliability criterion, the original relation predicate is determined as a new relation predicate to obtain standardized triple data based on the new relation predicate.

[0091] The technical solution of this application can automatically discover and register new relationships that meet the pre-set reliability standards by combining contextual information, realize the adaptive expansion of the knowledge graph, reduce the cost of manually maintaining the dynamic ontology library, and improve the ability to continuously evolve in scenarios with highly heterogeneous data such as finance and industry.

[0092] Optionally, see Figure 4 Step S400 may include steps S410 to S430.

[0093] In step S410, the standardized triplet data is entity aligned to obtain aligned standardized triplet data.

[0094] According to the example implementation, entity alignment can eliminate entity redundancy caused by multiple document sources and multiple representations, ensuring that the same entity has one and only one corresponding node in the knowledge graph.

[0095] For example, in step S410, the graph construction module 24 performs entity alignment on the standardized triplet data to obtain aligned standardized triplet data. The graph construction module 24 merges different representations (such as full name and abbreviation, Chinese name and English name) referring to the same entity by recognizing the semantic features of the entities, thus obtaining aligned standardized triplet data. This setup effectively removes redundant data, ensuring the uniqueness of entities in the knowledge graph and the connectivity of the knowledge network.

[0096] In step S420, the subjects and objects of the aligned standardized triplet data are mapped to nodes.

[0097] According to the example embodiment, in step S420, the graph construction module 24 maps the subjects and objects of the aligned standardized triplet data to nodes. The graph construction module 24 can perform a graph instantiation operation, mapping the subjects and objects of the aligned standardized triplet data to nodes according to the storage structure of the multimodal knowledge graph database.

[0098] In step S430, the relation predicates of the aligned and standardized triplet data are mapped to directed edges.

[0099] According to the example embodiment, in step S430, the graph construction module 24 maps the relation predicates of the aligned standardized triplet data to directed edges. The graph construction module 24 can perform a graph instantiation operation, mapping the relation predicates (i.e., standard relation predicates or new relation predicates) of the aligned standardized triplet data to directed edges based on the storage structure of the multimodal knowledge graph database.

[0100] The technical solution of this application aligns standardized triple data into entities to obtain aligned standardized triple data. The subjects and objects of the aligned standardized triple data are mapped to nodes, and the relation predicates of the aligned standardized triple data are mapped to directed edges. This technical solution effectively removes redundant data by aligning standardized triple data into entities, ensuring the uniqueness of entities in the knowledge graph and the connectivity of the knowledge network.

[0101] According to another aspect of this application, this application also provides a non-volatile computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is capable of implementing the knowledge graph construction method as described above.

[0102] According to another aspect of this application, this application also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the knowledge graph construction method as described above.

[0103] According to another aspect of this application, this application also provides a computer program product, including: a computer program stored on a computer-readable storage medium; the computer program includes program instructions that, when executed by a computer, cause the computer to perform the knowledge graph construction method as described above.

[0104] Finally, it should be noted that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions of the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing a knowledge graph, characterized in that, include: The received unstructured documents are decoupled into a streaming process to obtain a plain text data stream and a visual data stream. Semantic direct reading and extraction are performed on the plain text data stream and the visual data stream to obtain a triple pool; The triplet pool is then summarized and standardized to obtain standardized triplet data. The standardized triplet data is mapped to nodes and directed edges; Construct a knowledge graph based on the nodes and the directed edges.

2. The knowledge graph construction method according to claim 1, characterized in that, The step of performing semantic direct reading and extraction on the plain text data stream and the visual data stream to obtain a triple pool includes: The visual data stream is subjected to global perception description, and corresponding natural language description information is generated; The plain text data stream, the visual data stream, and the natural language description information are mixed and extracted to generate the triple pool.

3. The knowledge graph construction method according to claim 1, characterized in that, The process of summarizing and standardizing the triplet pool to obtain standardized triplet data includes: Determine whether the original relation predicate of the triple pool is a known relation predicate; If so, map the original relation predicate to a standard relation predicate to obtain the standardized triplet data based on the standard relation predicate; If not, determine whether the original relation predicate meets the preset reliability criteria; If the original relation predicate meets the preset reliability criteria, the original relation predicate is determined as the new relation predicate, so as to obtain the standardized triplet data based on the new relation predicate.

4. The knowledge graph construction method according to claim 3, characterized in that, The step of determining whether the original relation predicate meets the preset reliability criteria includes: Based on the context information of the original relation predicate in the unstructured document, determine whether the original relation predicate satisfies the preset reliability criteria.

5. The knowledge graph construction method according to claim 1, characterized in that, The step of mapping the standardized triplet data to nodes and directed edges includes: The standardized triplet data is entity aligned to obtain aligned standardized triplet data; Map the subject and object of the aligned, standardized triplet data to the nodes; The relational predicates of the aligned, standardized triplet data are mapped to the directed edges.

6. A knowledge graph construction system, characterized in that, include: The document parsing engine module performs streaming decoupling on the received unstructured documents to obtain plain text data streams and visual data streams; The semantic direct reading operator module performs semantic direct reading and extraction on the plain text data stream and the visual data stream to obtain a triple pool; An adaptive induction module performs induction and standardization processing on the triple pool to obtain standardized triple data; The graph construction module maps the standardized triplet data into nodes and directed edges, and constructs a knowledge graph based on the nodes and directed edges.

7. The knowledge graph construction system according to claim 6, characterized in that, The document parsing engine performs global perception and description on the visual data stream, and generates corresponding natural language description information; The document parsing engine performs mixed extraction of the plain text data stream, the visual data stream, and the natural language description information to generate the triple pool.

8. The knowledge graph construction system according to claim 6, characterized in that, The semantic direct reading operator module determines whether the original relation predicate of the triple pool is a known relation predicate; If so, the semantic direct reading operator module maps the original relation predicate to a standard relation predicate to obtain the standardized triplet data based on the standard relation predicate; If not, the semantic direct reading operator module determines whether the original relation predicate meets the preset reliability standard; If so, the semantic direct reading operator module determines the original relation predicate as a new relation predicate if the original relation predicate meets the preset reliability standard, so as to obtain the standardized triplet data based on the new relation predicate.

9. The knowledge graph construction system according to claim 8, characterized in that, The semantic direct reading operator module determines whether the original relation predicate satisfies the preset reliability standard based on the context information of the unstructured document.

10. The knowledge graph construction system according to claim 6, characterized in that, The graph construction module performs entity alignment on the standardized triplet data to obtain aligned standardized triplet data; The graph construction module maps the subject and object of the aligned and standardized triplet data to the nodes; The graph construction module maps the relation predicates of the aligned, standardized triplet data to the directed edges.

Images