Printing field knowledge graph automatic construction method based on large language model

Abstract

The invention discloses a printing field knowledge graph automatic construction method based on a large language model, which comprises the following steps: collecting a printing field multi-source heterogeneous document, processing the document to form a structured text block set, and taking the structured text block set as the input of the large language model; the model generates a primary abstract for each document, then performs cross-document semantic association analysis, constructs a document-level access sequence and a potential association network, and generates a secondary abstract; based on a printing field ontology mode, constructing a knowledge extraction mechanism fused with a structured prompt template, and guiding the model to perform structured knowledge extraction to form a candidate triple; and performing entity alignment, conflict resolution and multi-source knowledge fusion on the triple to form a structured knowledge graph. According to the method, end-to-end and automatic construction from the unstructured document to the structured knowledge graph is realized, the integrity, reasonability and construction efficiency of knowledge organization are remarkably improved, and core technical support is provided for printing intelligent diagnosis, process optimization and knowledge reuse.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of knowledge graph construction and natural language processing, and relates to a printing field knowledge graph automatic construction method based on a large language model. BACKGROUND

[0002] In the printing manufacturing field, enterprises accumulate a large amount of unstructured technical documents, fault records, process manuals and expert experience and other knowledge resources. These knowledge is usually scattered in the form of text, reports, charts and the like, and lacks a unified structured representation, resulting in the following problems: (1) knowledge fragmentation and islandization, printing knowledge is scattered in different systems, personnel and documents, and is difficult to integrate and share; (2) low efficiency of knowledge acquisition and utilization, relying on manual review and experience summary, slow response speed, and easily affected by subjective factors; (3) lack of structured association, there is a lack of explicit semantic association between documents, and it is difficult to form a knowledge network that can be reasoned and queried; (4) limitations of traditional knowledge extraction methods, methods based on rules or traditional machine learning rely on a large amount of labeled data and have poor generalization ability, and are difficult to adapt to the characteristics of texts in the printing field which are highly professional and diverse in expression.

[0003] At present, although there has been some progress in knowledge extraction methods based on large language models, there are still the following shortcomings: (1) most methods rely on pre-defined templates or structured data, and it is difficult to directly process non-normalized, multi-source heterogeneous texts commonly found in the printing field; (2) the extraction results are often isolated triples, and lack the ability to construct semantic associations at the document level and across documents; (3) in terms of professional terminology and complex causal relationships, existing methods lack accuracy and interpretability, and are difficult to meet the needs of industrial-level knowledge graph construction.

[0004] Therefore, there is an urgent need for a method that can automatically and accurately construct a knowledge graph from unstructured documents in the printing field to support subsequent applications such as intelligent retrieval, fault diagnosis and process optimization. SUMMARY

[0005] The purpose of the present application is to provide a printing field knowledge graph automatic construction method based on a large language model, which realizes the automatic, end-to-end and high-precision extraction of structured knowledge from unstructured documents in the printing field, constructs a queryable and reasonable knowledge graph, and significantly improves the completeness, reasonability and construction efficiency of knowledge organization.

[0006] The technical solution adopted by the present application is: The printing field knowledge graph automatic construction method based on a large language model, specifically comprises: S1, collecting multi-source heterogeneous documents in the printing field, processing the documents to form a set of structured text blocks, and taking the set of structured text blocks as the input of the large language model; S2, the large language model generates a primary summary for each document, then performs cross-document semantic association analysis, constructs a document-level access order and a potential association network, and generates a secondary summary set; S3, based on a predefined ontology model in the printing field, a knowledge extraction mechanism for a fusion structured prompt template in the printing field is constructed, the model is guided to perform structured knowledge extraction, and a candidate triple is formed; the candidate triple is subjected to entity alignment, conflict resolution and multi-source knowledge fusion, and a structured printing field knowledge graph is formed.

[0007] The characteristics of the present application also include: Step S1 is specifically: Collecting multi-source heterogeneous documents in the printing field, performing data cleaning and standardization processing on the documents, and then performing semantic blocking processing based on the document structure characteristics in the printing field to form a structured text block set; The specific steps of the semantic blocking processing based on the document structure characteristics in the printing field are: First, the inherent printing technology structural paragraph title in the document is recognized; second, the text is preliminarily divided at the natural paragraph boundary, and the theme vector of the continuous paragraph is calculated; finally, a hierarchical agglomerative clustering algorithm is applied, the structural paragraph title is used as a hard boundary, and the paragraph theme coherence is used as a soft constraint, and the continuous paragraphs with similar themes and no hard structural separation are combined into a semantically complete structured text block.

[0008] Step S2 is specifically: Constructing field-specific prompt words to guide the large language model to perform primary summary operation on flat text, and generating primary summaries for input scattered documents without obvious association relationship; cross-document semantic association analysis is performed according to the primary summary content, so as to construct a document-level access order and a potential association network, and generate a secondary summary set with cross-document semantic association.

[0009] The cross-document semantic association analysis algorithm in step S2 specifically includes: (1) Association reasoning prompt construction: design a hierarchical structure prompt word template to guide the large language model to identify the core printing field concepts that repeatedly appear in all primary summaries, and to judge the logical association relationship between any two summary contents; (2) Adaptive access order construction: calculate the semantic similarity and theme coherence between document summaries to guide the large language model to give the best processing order of the dynamic programming document, and form a globally optimal document reading path; (3) Secondary summary generation: According to the optimal processing order given in step (2) above, the document is processed in turn, and a context-enhanced prompt is constructed for the current document, which explicitly contains the core conclusions of the previous documents and the recognized associated relationships. The prompt instructs the large language model to generate a secondary summary that not only summarizes the content of the current document but also explicitly clarifies its knowledge association with the processed documents.

[0010] In step S3, the specific construction process of the predefined printing domain ontology model is as follows: Extract candidate terms from standard documents in the printing industry, and use a large language model to perform term emergence analysis on the general domain corpus to form an initial domain concept pool. For each candidate term in the concept pool, calculate its semantic similarity with known standard terms, and use term context feature analysis to merge it into the most relevant standardized parent concept to eliminate semantic ambiguity and automatically construct an initial ontology framework with a "class-subclass-instance" hierarchy. Finally, through a verification and correction cycle guided by expert knowledge, the initial framework is iteratively modified in multiple rounds in an interactive manner with experts to form the printing domain ontology model, clearly defining entity types and relationship types.

[0011] In step S3, the specific workflow of the knowledge extraction mechanism that fuses the structured prompt template is as follows: First, perform entity recognition and classification. The template instructs the large language model to identify all phrases in the text that belong to predefined entity types and classify them. Then, perform relationship context limited extraction. The template guides the large language model to extract relationships based on the identified entity pairs only under the condition that the predefined domain relationship context pattern is met. Finally, perform structured output. The template forces the large language model to output triples in a specified structured format. Before each call to the large language model for extraction, dynamic few-sample example selection and injection is performed: From the pre-built printing domain extraction sample library, according to the deep semantic similarity and entity type distribution matching degree between the current text block and the samples, 3-5 most relevant samples are retrieved in real time and inserted into the prompt template as "context learning" samples.

[0012] In step S3, entity alignment is as follows: Use a semantic embedding model to vectorize and hierarchically cluster all entity references, and then combine the pre-built printing standard term library and synonym rules to perform domain rule-guided forced merging on the clustering results, thereby assigning a unique standard identifier to each entity reference.

[0013] In step S3, conflict resolution is achieved through a multi-evidence weighted relationship conflict detection and arbitration algorithm, which is used to resolve contradictory descriptions of the same fact relationship in different documents. Specifically, it includes: (1) Multi-dimensional evidence collection and quantification: For the conflict relationship between the same pair of entities, the system extracts and quantifies two types of evidence from the original data: Source evidence weight: according to the different types of source documents, give its basic authority score; Domain logic consistency score: compare the relationship with the rules declared in the pre-defined ontology model of the printing field, and give a logical compatibility score; (2) Dynamic arbitration and decision: Design a weighted arbitration function, whose input is the score of the above two types of evidence, and the output is the comprehensive confidence of the relationship statement; Set a relative winning threshold, only when the confidence of one direction is significantly higher than that of the other, and the difference between the two exceeds the threshold, the relationship in this direction is adopted, otherwise it is marked as "to be reviewed".

[0014] The beneficial effects of the present application are: (1) End-to-end automation: without manual annotation of data or design of complex rules, the whole process of automatic construction from original unstructured documents to knowledge graph is realized, greatly reducing the labor cost and time cost; (2) Enhanced semantic association: by generating primary abstract, analyzing cross-document semantic association and planning the optimal processing order, and then generating secondary abstract with enhanced association, the cross-document knowledge association is strengthened, and the coherence and integrity of the knowledge graph are improved; (3) Good field adaptability: with the powerful semantic understanding ability of large language model, combined with the ontology model and structured prompt template of the printing field, it can effectively handle printing professional terms, complex sentence patterns and diversified expressions in the field, and adapt to the professional characteristics of the printing field; (4) Strong scalability: supporting dynamic addition of printing field documents and entity types, facilitating continuous updating and iterative optimization of knowledge graph according to industry development and business expansion, long-term adaptation to various application scenarios; (5) High knowledge quality: through entity alignment, conflict resolution and multi-source knowledge fusion mechanism, the problems of inconsistent entity expression and contradictory relationship are effectively solved, ensuring the accuracy, standardization and low redundancy of the knowledge graph. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the method of the present application; Figure 2 is a framework diagram of the method of the present application; Figure 3 is a knowledge graph construction flowchart in the scenario of embodiment 1 of the present application; Figure 4 is a knowledge graph construction flowchart in the scenario of embodiment 2 of the present application. DETAILED DESCRIPTION

[0016] The application will be described in detail below in conjunction with the drawings and specific embodiments.

[0017] The application is based on a large language model-based automatic construction method of a printing field knowledge graph, which is implemented according to the following steps as shown in Figure 1 and Figure 2 . S1, data preprocessing. It aims to convert the original heterogeneous printing field documents into standardized and pure text units, laying a foundation for subsequent processing. The specific process is as follows: S11, multi-source document input and analysis. Collecting multi-source documents such as printing equipment manufacturer manuals, enterprise production records, fault maintenance reports, and process standard files, etc. The formats include PDF, Word, TXT, etc. Use the corresponding analysis tools (such as PyPDF2, python-docx) to extract the pure text content in the document, and filter out non-text elements such as images, tables, and format labels.

[0018] S12, text cleaning and standardization. Perform cleaning operations on the extracted text, including: removing garbled codes, headers and footers, irrelevant symbols; uniformly standardizing domain-specific terms and abbreviations.

[0019] S13, semantic block processing based on printing field document structure features. To overcome the problem of destroying semantic integrity by traditional fixed-length segmentation, the application adopts a block algorithm based on the structure features of printing field documents. First, the algorithm identifies and uses the inherent printing technology structural paragraph titles in the document, such as “fault phenomenon”, “reason analysis”, “solution steps”, “parameter setting”, “notes”, etc. as hard boundaries for segmentation. Second, the cleaned text is segmented at the natural paragraph boundaries. Then, the theme vector of the continuous paragraph is calculated, and the hierarchical agglomerative clustering algorithm is applied. Within the framework of hard boundary formed by structural titles, the theme coherence of paragraphs is used as a soft constraint to merge continuous paragraphs with similar themes and no hard structural separation. Finally, a set of structured text blocks that fit the internal logic units of the printing field knowledge and are beneficial to the subsequent precise understanding and knowledge extraction of large language models is output, denoted as document set , as the input of the subsequent module.

[0020] S2, flat text summary and association construction. This step aims to construct a global semantic view and an optimal processing sequence for the document set that is still in a “flat” state and has no explicit association after preprocessing. The specific process is based on a cross-document semantic association analysis algorithm, which is as follows: S21, generating a primary summary. For each document all text blocks are concatenated and inputted into a large language model. By constructing domain-specific prompt words (e.g., "Please generate a refined summary for the following printing technology document, highlighting its core topics, key equipment involved, process points, or typical defects."), a primary summary is generated for each inputted scattered document without obvious association .

[0021] S22, correlation reasoning and adaptive access order construction. The primary summaries of all documents are inputted into a large language model again to perform correlation reasoning prompt construction. By designing a prompt word template with a hierarchical structure, the model is guided to identify the core printing field concepts that repeatedly appear in all primary summaries and to determine the logical correlation (such as causality, time sequence, reference) between any two summary contents. Based on this analysis, the model further calculates the semantic similarity and thematic coherence between document summaries, thereby dynamically planning a globally optimal document processing order list .

[0022] S23, generation of correlation-enhanced secondary summaries. According to the optimal processing order determined in step S22, the documents are processed in sequence. A context-enhanced prompt is constructed for the current document, which explicitly includes the core conclusions of the previous documents and the identified correlation. The large language model is instructed to generate a secondary summary that not only summarizes the current document content but also explicitly illustrates its correlation with the previously processed document knowledge (e.g., "Please integrate the current document content based on the analyzed document background, and generate a more detailed summary that reflects the correlation with the previously mentioned knowledge.") , providing a rich collaborative context for subsequent knowledge extraction.

[0023] S3, pattern-guided knowledge extraction and graph generation. Through precise extraction and fusion guided by domain knowledge, the construction of a knowledge graph is completed. The specific process is as follows: S31. Construct a predefined ontology model for the printing domain. First, perform multi-source standard terminology collection and candidate generation. Extract candidate terminology lists from ISO standards, national / industry technical specifications, technical white papers from mainstream equipment manufacturers, and authoritative process manuals in the printing industry using rules and statistical methods to form an initial domain concept pool. Then, use a large language model to perform term emergence analysis on the general corpus of the printing domain, identifying potential technical concepts that are not covered by standard documents but frequently co-occur, and adding them to the concept pool. Second, perform semantic disambiguation and hierarchical structuring: for each candidate term in the concept pool, calculate its semantic similarity to known standard terms, and use terminology context feature analysis to merge them under the most relevant standardized parent concept to eliminate semantic ambiguity and automatically construct an initial ontology framework with a "class-subclass-instance" hierarchy. Finally, present the initial ontology framework to domain experts in an interactive and visual manner, allowing experts to confirm, split, or adjust the automatic merging results. After each modification, through 3-5 rounds of rapid iteration, a stable, authoritative, and expert-consensus ontology model for the printing field is formed, which includes a set of strictly defined entity types (such as defect phenomena, equipment parts, process parameters, and materials) and relationship types (such as causing, occurring at, being affected by, and belonging to).

[0024] S32, Knowledge Extraction from Structured Prompt Templates. Based on the above ontology model, a prompt template with a strict logical hierarchical structure is constructed. This template instruction big language model first identifies and categorizes phrases in the text belonging to predefined entity types (such as printing defects, equipment parts, process parameters); then, based on the identified entity pairs, it extracts relations only under contextual patterns that satisfy predefined domain relationships (such as causing, occurring at, affected by); finally, it forces the output of triples in a specified structured format (such as JSON). Before each extraction, dynamic few-shot example selection and injection are performed: from a high-quality printing domain sample library, based on the deep semantic similarity and entity type distribution matching degree between the current text block and the sample, the 3-5 most relevant examples are retrieved in real time. These examples, in their original "text-structured output" pairing format, are dynamically inserted after the system instruction part of the hierarchical prompt template as "context learning" samples for the big language model, thereby significantly improving the model's extraction accuracy for specific fault modes or process scenarios. Combined with the secondary summary provided in step S2... As a context, the document content is combined with the patterned prompt template and input into the large language model to extract a structured preliminary list of candidate triples.

[0025] S33, Entity Alignment, Conflict Resolution, and Multi-Source Knowledge Fusion. This step aims to integrate and refine the extracted candidate triples to form a unified and standardized knowledge set.

[0026] Entity alignment is achieved through domain knowledge-enhanced semantic alignment and normalization. Specifically, it includes: First, for all entities extracted from multiple documents, using semantic embedding models such as BGE, calculating the vector representations of all entity references and their pairwise cosine similarities to perform preliminary hierarchical semantic clustering; Second, performing domain rule-guided forced merging. If any term within a cluster matches an entry in a printed standard terminology library, that standard term is set as the standard entity of that cluster. Simultaneously, if terms within a cluster can be associated with the same standard term through synonym rules, forced merging is performed; Finally, each original entity reference is assigned a unique, standardized standard entity identifier, completing entity alignment.

[0027] Conflict resolution is achieved through a multi-evidence weighted relation conflict detection and arbitration algorithm, used to resolve contradictory descriptions of the same factual relationship in different documents, specifically including: 1) Multi-dimensional evidence collection and quantification: For each pair of conflicting relationships, the system extracts and quantifies two types of evidence from the raw data: Source evidence weighting: A basic authority score is assigned based on the different types of source documents (e.g., national standards are weighted at 1.0, company manuals at 0.7, and personal repair records at 0.4). Domain logical consistency score: Compare the relationship with the rules declared in the predefined printing domain ontology pattern (such as "ink imbalance may lead to dirty plates, but the reverse is not true") and give a logical compatibility score; 2) Dynamic Arbitration and Decision-Making: Design a weighted arbitration function whose input is the scores of the two types of evidence mentioned above, and whose output is the overall confidence level of the relationship statement. For conflicting relationship pairs, calculate the overall confidence level for each direction. Set a relative winning threshold (e.g., 20%): only when the confidence level of one direction is significantly higher than that of the other (exceeding the threshold) is the relationship in that direction adopted; if the difference in confidence levels between the two parties is less than the threshold, the conflict is marked as "pending review," and all its evidence summaries are stored in the pending queue for subsequent manual or more complex automated reasoning intervention.

[0028] S34, Graph Storage and Output. The final processed triplet data is imported into a graph database such as Neo4j for storage. Standard entities serve as nodes, relationships as edges, and related attributes can be attached. The final result is a queryable, displayable, and applicable structured knowledge graph for the printing field, suitable for intelligent diagnostics and process optimization.

[0029] Example 1: This embodiment illustrates a method for constructing a defect diagnosis knowledge graph based on fault repair reports. The method of this invention addresses the following scenario: A printing company wants to construct a knowledge graph from its accumulated 100 historical fault repair reports to quickly locate the root causes of common faults. For example... Figure 3 As shown, the specific implementation process is as follows: S1: Data Preprocessing S11: Collect 100 PDF "Fault Repair Reports", use a PDF parsing library to extract the plain text, and remove irrelevant elements such as signatures and tables.

[0030] S12: Clean the text, standardize terminology (e.g., unify "ghosting" to "double image"), and divide the text into blocks according to the natural paragraphs of the report.

[0031] S13: Output the cleaned text block set. This step applies a block segmentation algorithm based on the document structure characteristics of the printing industry. By identifying the inherent title structure in the report, such as "fault phenomenon" and "cause analysis", and combining paragraph topic clustering, the report content is divided into semantically complete and logically coherent text blocks, totaling about 1200 blocks, which constitute the document set to be processed.

[0032] S2: Flat Text Summarization and Association Building S21: Use the large language model to generate a preliminary summary for each report. For example, the summary for report #45 is: "This report describes a case of 'dirty printing' caused by 'ink-water imbalance,' and records the troubleshooting process and the final adjusted process parameters." S22: Through cross-document semantic association analysis algorithm, analyze the primary summary of all reports, identify the distribution of knowledge topics and potential associations by large language model, and dynamically plan the processing order from "basic principle type" reports to "complex comprehensive fault type" reports.

[0033] S23: Following the above order, guide the large language model to generate secondary summaries, strengthening knowledge connections across reports. For example, when processing subsequent reports, the summaries will refer to the "ink-water balance" concept and other fault variants that appeared in previous reports.

[0034] S3: Pattern-Guided Knowledge Extraction and Graph Generation S31: Based on a pre-built ontology model of the printing domain (which is formed by integrating industry standard terminology and being verified by experts), construct a prompt template to guide the large language model in structured extraction.

[0035] S32: Extract preliminary triplets from each report. For example, extract from report #45: ["Dirty printing", "Causes", "Ink-water imbalance"], ["Ink-water imbalance", "Related parameters", "Water supply"].

[0036] S33: Perform entity alignment and fusion for triples in all reports. For example, the system aligns the terms "ghosting" and "double image" from different reports to the standard entity "misregistration". The system found "ink dot" defects in 20 reports, 15 of which pointed to "high ink viscosity". Based on a multi-evidence weighted arbitration algorithm, the system adopted this as the primary relation and retained and labeled "environmental dust" as a secondary cause.

[0037] S34: Import the final triples into the Neo4j graph database to generate a fault diagnosis knowledge graph containing more than 500 nodes and more than 800 relationships, supporting source tracing queries starting from defect phenomena.

[0038] Example 2: This embodiment illustrates a method for constructing a process parameter knowledge graph by integrating operation manuals from multiple printing press models. The scenario is as follows: A printing equipment manufacturer wants to integrate the operation manuals and process parameter guidelines for multiple models to build a unified knowledge base to support intelligent customer service. For example... Figure 4 As shown, the specific implementation process is as follows: S1: Data Preprocessing S11: Collect structured Word manuals and unstructured PDF parameter guides for various models, parse and extract the text content, and convert the parameter tables in the manuals into descriptive text of "parameter name: value, description".

[0039] S12: Clean up the text and unify the different descriptions of the same operation in various manuals (e.g., unify "start-up self-test" and "power-on self-test" into "device self-test").

[0040] S13: Output a standardized set of text blocks. Semantic block processing was completed by identifying and analyzing the titles of sections such as "Operating Steps" and "Parameter Settings" in the user manual.

[0041] S2: Flat Text Summarization and Association Building S21: Generate a preliminary summary for each manual and guide, outlining its core content, such as "Model Alpha Manual: Covers operating procedures such as startup, printing pressure settings, and routine maintenance." S22: The large language model analyzes all summaries, identifies "basic operations" as shared modules and "advanced functions" as model-specific modules, and plans the processing order from general to specific based on this.

[0042] S23: Generate secondary summaries in sequence, clearly indicating the specific machine model to which each operation step or parameter is applicable, and establish cross-document process associations.

[0043] S3: Pattern-Guided Knowledge Extraction and Graph Generation S31: Defines a domain pattern that includes entities such as equipment model, operating steps, process parameters, and target effects.

[0044] S32: Use patterned prompts for extraction. For example, from the text “The recommended printing pressure for model Alpha is 200±10N to ensure clear dots”, extract: [“model Alpha”, “with parameters”, “printing pressure”], [“printing pressure”, “recommended value”, “200±10N”], [“printing pressure (200±10N)”, “target effect”, “clear dots”].

[0045] S33: Merge different recommended values ​​for the same parameter (such as "printing speed") from different manuals. The system first aligns and confirms the entity "printing speed", then applies conflict resolution strategies based on the authority of the source manuals (such as OEM manuals vs. general guidelines), and finally adds attributes such as "applicable models" and "value range" to the parameter node in the knowledge graph to achieve standardized integration of knowledge.

[0046] S34: Construct a cross-model process knowledge graph. Based on this graph, the customer service system can accurately answer users' questions about parameter differences between different models under specific process objectives.

[0047] Example 3: This embodiment presents an automated method for constructing a knowledge graph in the printing domain based on a large language model, specifically as follows: S1 collects heterogeneous documents from multiple sources in the printing field, processes the documents to form a structured set of text blocks, and uses it as input for a large language model. S2: The large language model generates a primary summary for each document, then performs cross-document semantic association analysis, constructs a document-level access order and potential association network, and generates a set of secondary summaries; S3, based on a predefined ontology model for the printing domain, constructs a knowledge extraction mechanism for the printing domain that integrates structured prompt templates, guides the model to extract structured knowledge, and forms candidate triples; entity alignment, conflict resolution and multi-source knowledge fusion are performed on the candidate triples to form a structured knowledge graph for the printing domain.

[0048] Example 4: Based on Example 3, step S1 specifically includes: Collect heterogeneous documents from multiple sources in the printing field, clean and standardize the documents, and then perform semantic segmentation based on the structural features of the documents in the printing field to form a set of structured text blocks. The specific steps of semantic segmentation processing based on the document structure features in the printing field are as follows: First, identify the inherent typographical structure paragraph headings in the document; second, perform initial segmentation of the text at natural paragraph boundaries and calculate the topic vectors of consecutive paragraphs; finally, apply a hierarchical agglomerative clustering algorithm, using the aforementioned structure paragraph headings as hard boundaries and combining paragraph topic coherence as a soft constraint, to merge consecutive paragraphs with similar topics and no hard structural separations into semantically complete structured text blocks.

[0049] Example 5: Based on Example 4, step S2 specifically includes: Domain-specific cue words are constructed to guide a large language model to perform a flat text primary summarization operation, generating primary summaries from scattered input documents with no obvious correlation. Based on the content of the primary summaries, cross-document semantic association analysis is performed to construct a document-level access order and potential association network, generating a set of secondary summaries with cross-document semantic associations.

[0050] Example 6: Based on Example 5, the cross-document semantic association analysis algorithm in step S2 specifically includes: (1) Construction of related reasoning prompts: Design prompt word templates with hierarchical structure to guide the large language model to identify core print domain concepts that appear repeatedly in all primary summaries, and at the same time determine the logical relationship between any two summaries; (2) Adaptive access order construction: Calculate the semantic similarity and topic coherence between document summaries, guide the large language model to give the optimal processing order of documents through dynamic programming, and form the globally optimal document reading path; (3) Secondary summary generation: Based on the optimal processing order given in step (2) above, the documents are processed sequentially, and context-enhanced hints are constructed for the current document. These hints clearly contain the core conclusions of the previous documents and the identified relationships. The instruction big language model generates a secondary summary that summarizes the content of the current document and explicitly explains its relationship with the knowledge of the processed documents.

Claims

1. An automated method for constructing knowledge graphs in the printing domain based on a large language model, characterized in that, Specifically: S1 collects heterogeneous documents from multiple sources in the printing field, processes the documents to form a structured set of text blocks, and uses it as input for a large language model. S2: The large language model generates a primary summary for each document, then performs cross-document semantic association analysis, constructs a document-level access order and potential association network, and generates a set of secondary summaries; S3, based on a predefined ontology model for the printing domain, constructs a knowledge extraction mechanism for integrated structured prompt templates for the printing domain, guiding the model to extract structured knowledge and form candidate triples; Entity alignment, conflict resolution, and multi-source knowledge fusion are performed on candidate triples to form a structured knowledge graph for the printing domain.

2. The method for automated construction of a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, Step S1 is as follows: Collect heterogeneous documents from multiple sources in the printing field, clean and standardize the documents, and then perform semantic segmentation based on the structural features of the documents in the printing field to form a set of structured text blocks. The specific steps of semantic segmentation processing based on the document structure features in the printing field are as follows: First, identify the inherent typographical structure paragraph headings in the document; second, perform initial segmentation of the text at natural paragraph boundaries and calculate the topic vectors of consecutive paragraphs; finally, apply a hierarchical agglomerative clustering algorithm, using the aforementioned structure paragraph headings as hard boundaries and combining paragraph topic coherence as a soft constraint, to merge consecutive paragraphs with similar topics and no hard structural separations into semantically complete structured text blocks.

3. The method for automated construction of a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, Step S2 is as follows: Domain-specific cue words are constructed to guide a large language model to perform a flat text primary summarization operation, generating primary summaries from scattered input documents with no obvious correlation. Based on the content of the primary summaries, cross-document semantic association analysis is performed to construct a document-level access order and potential association network, generating a set of secondary summaries with cross-document semantic associations.

4. The method for automated construction of a knowledge graph in the printing domain based on a large language model according to claim 1 or 3, characterized in that, The cross-document semantic association analysis algorithm in step S2 specifically includes: (1) Construction of related reasoning prompts: Design prompt word templates with hierarchical structure to guide the large language model to identify core print domain concepts that appear repeatedly in all primary abstracts, and at the same time determine the logical relationship between any two abstract contents; (2) Adaptive access order construction: Calculate the semantic similarity and topic coherence between document summaries, guide the large language model to give the optimal processing order of documents through dynamic programming, and form the globally optimal document reading path; (3) Secondary summary generation: Based on the optimal processing order given in step (2) above, the documents are processed sequentially, and context-enhanced hints are constructed for the current document. These hints clearly contain the core conclusions of the previous documents and the identified relationships. The instruction big language model generates a secondary summary that summarizes the content of the current document and explicitly explains its relationship with the knowledge of the processed documents.

5. The method for automated construction of a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, In step S3, the specific construction process of the predefined printing domain ontology pattern is as follows: Candidate terms are extracted from standard documents in the printing industry, and emergent terminology analysis is performed on the domain-wide corpus using a large language model to form an initial domain concept pool. For each candidate term in the concept pool, its semantic similarity with known standard terms is calculated, and with the help of terminology context feature analysis, it is merged into the most relevant standardized parent concept to eliminate semantic ambiguity and automatically construct an initial ontology framework with a "class-subclass-instance" hierarchy. Finally, through an expert-guided verification and correction loop, experts are collaboratively revised in an interactive manner to perform multiple rounds of iterative corrections on the initial framework, forming a printing domain ontology model and clarifying entity types and relationship types.

6. The method for automated construction of a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, In step S3, the specific workflow of the knowledge extraction mechanism that integrates structured prompt templates is as follows: First, entity recognition and classification are performed. The template instructs the language model to identify all phrases in the text that belong to predefined entity types and classify and label them. Then, relation context-limited extraction is performed. The template guides the language model to extract relations only based on the identified entity pairs and in accordance with the predefined domain relation context pattern. Finally, structured output is performed. The template forces the language model to output triples in a specified structured format. Before each call to the large language model for extraction, dynamic few-sample selection and injection is performed: from the pre-built sample library of the printing domain, based on the deep semantic similarity and entity type distribution matching degree between the current text block and the sample, the 3-5 most relevant examples are retrieved in real time and used as "context learning" samples to dynamically insert prompt templates.

7. The method for automatically constructing a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, In step S3, entity alignment specifically involves: The semantic embedding model is used to vectorize and hierarchically cluster all entity references. Then, combined with a pre-built standard printed terminology library and synonym rules, the clustering results are subject to domain rule-guided forced merging, thereby assigning a unique standard identifier to each entity reference.

8. The method for automatically constructing a knowledge graph in the printing domain based on a large language model according to claim 1, characterized in that, In step S3, conflict resolution is achieved through a multi-evidence weighted relation conflict detection and arbitration algorithm, used to resolve contradictory descriptions of the same factual relationship in different documents, specifically including: (1) Multidimensional evidence collection and quantification: Regarding the conflict between the same pair of entities, the system extracts and quantifies two types of evidence from the raw data: Source evidence weighting: Assigning a basic authority score to the source document based on its type; Domain logical consistency score: The relation is compared with the rules declared in the predefined print domain ontology schema, and a logical compatibility score is given. (2) Dynamic arbitration and decision-making: Design a weighted arbitration function whose input is the scores of the two types of evidence mentioned above, and whose output is the overall confidence level of the relationship statement; set a relative winning threshold, and only when the confidence level of one side is significantly higher than that of the other side, and the difference between the two exceeds the threshold, is the relationship of this direction adopted, otherwise it is marked as "pending review".

Images