Chemical material field knowledge graph construction method and device based on inference large model prompt instruction

Abstract

The invention discloses a chemical material field knowledge graph construction method and device based on an inference large model prompt instruction. The method comprises the following steps: acquiring basic data information and experimental chart data; establishing a domain rule base and a knowledge graph basic framework document; establishing structured JSON data; constructing a Neo4j database according to the structured JSON data and a field rule base to perform structured data quality verification and dynamic enhancement so as to obtain the updated field rule base and the verified structured data; acquiring preprocessed atlas data according to the verified structured data; acquiring a multi-modal fusion preliminary map according to the preprocessed map data; and establishing a knowledge graph containing a reasoning result according to the multi-modal fusion preliminary graph. According to the method, the logical relationship in the article content can be better extracted, so that the content and the relationship of the established knowledge graph are more comprehensive.

Description

Technical Field

[0001] The present application relates to the field of knowledge graph technology, and specifically to a method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction and a device for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction. Background Art

[0002] In recent years, chemical materials science, a key supporting discipline for the development of new materials, new energy, and high-performance devices, has seen explosive growth in research results. This has been accompanied by a vast amount of unstructured technical literature, scientific papers, patent specifications, and standard documents. These texts contain a wealth of specialized knowledge, such as material names, composition structures, preparation processes, physical and chemical properties, and application scenarios. However, efficiently and accurately extracting valuable information from this complex, multi-source text and constructing it into a structured, searchable, and reasonable knowledge graph has become a major challenge in intelligent materials research and information management.

[0003] Traditional methods for constructing chemical materials knowledge graphs primarily rely on manual annotation, rule extraction, or traditional machine learning. While these methods can achieve a certain degree of accuracy initially, they often suffer from numerous limitations. On the one hand, manual annotation is time-consuming and labor-intensive, making it difficult to cope with the ever-increasing volume of data. On the other hand, extraction methods based on rules or shallow models lack semantic understanding and are unable to cope with the terminological ambiguity, implicit relationship expressions, and complex sentence structures common in chemical texts. This results in coarse granularity, narrow coverage, and poor semantic consistency in the constructed results.

[0004] In recent years, large language models based on deep learning (such as the GPT series, BERT, T5, etc.) have demonstrated excellent language understanding and generation capabilities in natural language processing tasks. Especially in open-domain information extraction, question-answering systems, summary generation and other tasks, their versatility and generalization capabilities have greatly surpassed traditional models. Despite this, directly applying these models to the highly professional field of chemical materials still faces challenges, such as insufficient understanding of domain terminology, unclear knowledge boundaries, and uncontrollable output. To improve the model's domain adaptability, researchers have begun to try to combine prompt instruction technology to guide large models to complete specified information extraction tasks with higher accuracy and stability.

[0005] However, in the specialized field of chemical materials, a systematic, automated, and task-oriented large-model-assisted knowledge graph construction method is currently lacking. Therefore, there is an urgent need to develop a new construction method that combines large language models with prompt instructions to improve the automation and professional accuracy of chemical materials knowledge extraction. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for constructing a knowledge graph in the field of chemical materials based on inference large model prompt instructions to solve at least one of the above-mentioned technical problems.

[0007] One aspect of the present invention provides a method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction, the method comprising:

[0008] Obtaining basic data information and experimental chart data, wherein the basic data information includes professional literature in the chemical industry and a field terminology system;

[0009] Establish a domain rule base and a knowledge graph basic framework document based on the basic data information;

[0010] Creating structured JSON data based on the basic data information and the domain rule library, wherein the structured JSON data includes entities, relationships, and attributes;

[0011] Building a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data;

[0012] Mapping and preprocessing the verified structured data to graphs to obtain preprocessed graph data;

[0013] Perform multimodal data fusion and association on the preprocessed atlas data and experimental chart data to obtain a preliminary multimodal fusion atlas;

[0014] Establish domain knowledge reasoning association data for the multimodal fusion preliminary graph, thereby obtaining a knowledge graph containing reasoning results.

[0015] Optionally, establishing a domain rule base and a knowledge graph basic framework document based on the basic data information includes:

[0016] Obtain original document collection;

[0017] Perform text preprocessing on the original document collection to obtain text data that can be used for recognition;

[0018] Extract features from text data that can be used for recognition to obtain core feature data;

[0019] Performing a logical chain construction on the core feature data to obtain logical chain data; wherein the logical chain data and the core feature data constitute structured document data;

[0020] Define node types, relationship types, and attribute rules for structured document data to obtain knowledge graph framework information;

[0021] Generate rule template library based on domain terminology system;

[0022] A rule priority table is generated according to the rule template library; the rule template library and the rule priority table constitute a domain rule library.

[0023] Optionally, the creating structured JSON data according to the basic data information and the domain rule library includes:

[0024] Prompt word design, rule embedding design, format instruction design, and entity type design are performed according to the basic data information and the domain rule library, thereby generating a structured instruction template;

[0025] Decomposing the task according to the basic data information and the structured instruction template to obtain structured semantic units;

[0026] The structured semantic unit is converted into JSON format, thereby obtaining structured JSON data.

[0027] Optionally, the step of constructing a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data includes:

[0028] Performing an integrity check on the structured JSON data, thereby obtaining an integrity check report and updated JSON data;

[0029] Perform domain rule filtering based on the updated JSON data and the domain rule library to obtain filtered data and rule triggering records;

[0030] Performing cross-validation enhancement based on the filtered data and the domain rule library to obtain cross-validation results and enhanced data;

[0031] The domain rule base is updated according to the enhanced data, thereby obtaining an updated domain rule base.

[0032] Optionally, mapping and preprocessing the verified structured data to a graph to obtain preprocessed graph data includes:

[0033] Obtaining a graph mapping rule table and preliminary mapped graph data based on the verified structured data;

[0034] Generate standardized atlas data and preprocessing logs based on the initially mapped atlas data;

[0035] Conflict detection and resolution are performed based on the standardized atlas data to obtain preprocessed atlas data.

[0036] Optionally, performing multimodal data fusion and association on the preprocessed atlas data and the experimental chart data to obtain a preliminary multimodal fusion atlas includes:

[0037] Processing the experimental chart data to obtain structured chart data;

[0038] Performing cross-modal association on the structured chart data and the pre-processed atlas data, thereby obtaining cross-modal association data;

[0039] A modality fusion atlas is obtained according to the cross-modal association data and the preprocessed atlas data.

[0040] Optionally, establishing domain knowledge reasoning association data for the multimodal fusion preliminary graph to obtain a knowledge graph containing reasoning results includes:

[0041] generating an enhanced knowledge graph based on the multimodal fusion preliminary graph;

[0042] Construct a knowledge graph containing reasoning results based on the enhanced knowledge graph.

[0043] Optionally, generating an enhanced knowledge graph based on the multimodal fusion preliminary graph includes:

[0044] Using the breadth-first search algorithm, we randomly select an entity node from the preliminary multimodal fusion graph as the starting point, traverse its directly related nodes and indirectly related nodes layer by layer, and record all entity-relationship combinations on the traversal path;

[0045] Count the number of occurrences of entity-relationship combinations in all traversal paths, and obtain a list of high-frequency association patterns for each entity-relationship combination whose occurrence number exceeds a preset threshold;

[0046] Perform causal inference on the list of high-frequency association patterns to obtain a list of causal relationship candidates;

[0047] The multimodal fusion preliminary graph is updated according to the causal relationship candidate list and the multimodal fusion preliminary graph to form an enhanced knowledge graph, wherein the enhanced knowledge graph includes the multimodal fusion preliminary graph and new causal relationship edges generated by the causal relationship candidate list.

[0048] Optionally, constructing a knowledge graph containing reasoning results based on the enhanced knowledge graph includes:

[0049] Import the enhanced knowledge graph into the graph database, preserving the original document source and reasoning path metadata;

[0050] Merge entity nodes with the same name to generate a knowledge graph containing reasoning results.

[0051] The present application also provides a device for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction, characterized in that the device for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction includes:

[0052] A basic data acquisition module, which is used to acquire basic data information and experimental chart data, wherein the basic data information includes professional literature in the chemical industry and a field terminology system;

[0053] An establishment module, wherein the establishment module is used to establish a domain rule library and a knowledge graph basic framework document based on the basic data information;

[0054] A structured JSON data acquisition module, wherein the structured JSON data acquisition module is used to establish structured JSON data according to the basic data information and the domain rule library, wherein the structured JSON data includes entities, relationships and attributes;

[0055] A quality verification module, which is used to construct a Neo4j database based on the structured JSON data and the domain rule library to perform quality verification and dynamic enhancement of the structured data, thereby obtaining an updated domain rule library and verified structured data;

[0056] A mapping module, which is used to map and preprocess the verified structured data from the structured data to the graph, thereby obtaining preprocessed graph data;

[0057] An association module is used to perform multimodal data fusion and association on the pre-processed atlas data and experimental chart data, thereby obtaining a preliminary multimodal fusion atlas;

[0058] The reasoning association module is used to establish domain knowledge reasoning association data for the multimodal fusion preliminary graph, thereby obtaining a knowledge graph containing reasoning results.

[0059] Beneficial effects

[0060] The present invention discloses a method for constructing a knowledge graph in the field of chemical materials based on prompt instructions of a large inference model. The core of the method lies in the deep integration of a large language model and prompt engineering technology, innovatively simulating the "thinking chain" reasoning path of scientific researchers in literature reading, and breaking through the technical bottlenecks in the existing knowledge graph construction process in terms of entity recognition accuracy, semantic understanding depth, relationship modeling granularity and structured output standardization, ensuring that the logical relationships in the article content can be better extracted, making the established knowledge graph content and relationships more comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to an embodiment of the present application;

[0062] Figure 2 This is a schematic diagram of the knowledge graph framework information of this application;

[0063] Figure 3 This is the knowledge graph construction process based on the large model prompting project in the second embodiment of this application;

[0064] Figure 4 This is a partial content display of the knowledge graph in the second embodiment of this application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the implementation of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below in conjunction with the drawings in the embodiments of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described embodiments are part of the embodiments of this application, not all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to be used to explain this application, and should not be understood as limitations on this application. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The embodiments of this application are described in detail below in conjunction with the drawings.

[0066] like Figure 1 The method for constructing a knowledge graph in the field of chemical materials based on the inference large model prompt instruction shown includes:

[0067] Obtaining basic data information and experimental chart data, wherein the basic data information includes professional literature in the chemical industry and a field terminology system;

[0068] Establish a domain rule base and a knowledge graph basic framework document based on the basic data information;

[0069] Creating structured JSON data based on the basic data information and the domain rule library, wherein the structured JSON data includes entities, relationships, and attributes;

[0070] Building a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data;

[0071] Mapping and preprocessing the verified structured data to graphs to obtain preprocessed graph data;

[0072] Perform multimodal data fusion and association on the preprocessed atlas data and experimental chart data to obtain a preliminary multimodal fusion atlas;

[0073] Establish domain knowledge reasoning association data for the multimodal fusion preliminary graph, thereby obtaining a knowledge graph containing reasoning results.

[0074] In this embodiment, establishing a domain rule base and a knowledge graph basic framework document based on the basic data information includes:

[0075] Obtain original document collection;

[0076] Perform text preprocessing on the original document collection to obtain text data that can be used for recognition;

[0077] Extract features from text data that can be used for recognition to obtain core feature data;

[0078] Performing a logical chain construction on the core feature data to obtain logical chain data; wherein the logical chain data and the core feature data constitute structured document data;

[0079] Define node types, relationship types, and attribute rules for structured document data to obtain knowledge graph framework information;

[0080] Generate rule template library based on domain terminology system;

[0081] A rule priority table is generated according to the rule template library; the rule template library and the rule priority table constitute a domain rule library.

[0082] The following is a further detailed explanation of the application's establishment of a domain rule base and a knowledge graph basic framework document based on the basic data information by way of examples. It can be understood that this example does not constitute any limitation to the application.

[0083] In this embodiment, the original document collection includes various papers, patents, reports, etc., and the format of the original document collection may be PDF format or WORD format.

[0084] First, the original document set must be preprocessed. For example, if it is in PDF format, a PDF parsing tool can be used to extract the text content and retain the paragraph structure. The text can be segmented into sentences (split by periods and line breaks) and words (marked with parts of speech, such as nouns and verbs) through natural language processing technology. Based on the domain dictionary (such as "hydroformylation" and "catalyst"), the terms are annotated using keyword matching methods.

[0085] It is understandable that if the format is WORD, the text can also be processed using existing technologies.

[0086] In the above manner, text data that can be used for recognition can be obtained.

[0087] In this embodiment, according to the predefined module template (such as "research background usually contains words such as 'in recent years' and 'with'"), each relevant paragraph of text data that can be used for identification is located through keyword and pattern matching. For example, to match the "research background" paragraph: look for sentences containing words such as "in recent years" or "with".

[0088] Perform semantic analysis on the matching paragraphs, identify the key components in the sentences, and extract information such as "industry needs" (such as "improving reaction efficiency") and "scientific research challenges" (such as "insufficient catalyst stability"), so as to obtain core feature data and labels corresponding to the core feature data. For example, labels may include: research background, key issues: the main scientific research issues raised in the literature, technological innovation, material system: the materials or systems described in the literature, experimental methods, performance data, etc.

[0089] Each core feature data is processed as follows:

[0090] Use syntactic analysis tools to parse sentence structures (such as identifying the “drive” relationship: the subject is “industry demand,” the predicate is “drive,” and the object is “scientific research challenge”).

[0091] According to the predefined logical chain relationship template (such as "industry demand → driver → scientific research challenge → solution → technological innovation"), a logical chain is constructed through rule matching (such as "in order to solve the problem..., this study proposes...") or manual annotation to obtain logical chain data.

[0092] In this embodiment, a complete structured document data can be as shown in Table 1:

[0093]

[0094]

[0095] In this embodiment, the structured document data is processed as follows:

[0096] Node type definition:

[0097] Entity Recognition:

[0098] Entities are manually or semi-automatically identified based on domain knowledge (e.g., “rhodium-based catalyst,” “impregnation method,” “TOF value”).

[0099] Type mapping:

[0100] Define node type labels (such as "catalyst", "preparation method", "performance index");

[0101] Map entities to predefined node types (e.g., "rhodium-based catalyst" is classified as "catalyst");

[0102] Output: Node type mapping table (such as "rhodium-based catalyst → catalyst, impregnation method → ​​preparation method").

[0103] Relationship type definition:

[0104] Relationship template library:

[0105] Define relationship templates (such as "drive", "resolve", "include", "promote").

[0106] Relationship Supplement:

[0107] Automatically supplement or manually define relationship types based on logical chain analysis (e.g., extracting the "driving" relationship from "industry demand → driving → scientific research challenge");

[0108] Output: List of relationship types (e.g. "Drives, Resolves, Contains, Promotes").

[0109] Attribute rule formulation:

[0110] Input: core module data (such as "catalyst particle size 5nm", "calcination temperature 500℃").

[0111] Attribute extraction:

[0112] Extract entity attributes (such as "particle size", "temperature") from the literature.

[0113] Rules and constraints:

[0114] Define numerical attribute rules based on domain knowledge (such as "particle size: unit nm, range 1-100").

[0115] Define classification attribute rules (such as "catalyst type: rhodium-based, cobalt-based, iron-based").

[0116] Output: attribute rule document (such as "particle size: unit nm, range 1-100; TOF value: unit h -1 , range 0-10000”).

[0117] Through the above processing, the knowledge graph framework information (such as table or document format) is obtained, and the node type, relationship type and attribute rules are defined. Figure 2 shown.

[0118] In this embodiment, the domain rule base is obtained in the following way:

[0119] Rule template conversion:

[0120] Numerical rules:

[0121] Convert terms into numerical constraints (e.g., "particle size > 100 nm" is converted to "if the particle size exceeds 100 nm, mark it as abnormal data").

[0122] Logical rules:

[0123] Convert process logic into rules (e.g., “impregnation must be completed before roasting” into “if the preparation method is impregnation and the roasting sequence is after impregnation, trigger a warning”).

[0124] Output: rule template library (such as "numerical rule: particle size > 100nm is marked as abnormal; logical rule: calcination after impregnation triggers a warning").

[0125] Rule priority assignment:

[0126] Input: rule template library;

[0127] Conflict Analysis:

[0128] Analyze potential conflicts between rules (e.g., “particle size range” vs. “the literature claims that the particle size is reasonable”).

[0129] Priority setting:

[0130] Setting priorities based on domain knowledge (e.g., the “sequence of preparation processes” rule takes precedence over the “reasonableness of performance values” rule);

[0131] Output: Rule priority table (such as "particle size rule priority 2, temperature rule priority 1, preparation process rule priority 3").

[0132] In this embodiment, the domain rule library (eg, in table or document format) includes rule templates, priorities, and conflict resolution strategies (as shown in Table 2 below).

[0133]

[0134]

[0135] In this embodiment, establishing structured JSON data based on the basic data information and the domain rule library includes:

[0136] Prompt word design, rule embedding design, format instruction design, and entity type design are performed according to the basic data information and the domain rule library, thereby generating a structured instruction template;

[0137] Decomposing the task according to the basic data information and the structured instruction template to obtain structured semantic units;

[0138] The structured semantic unit is converted into JSON format, thereby obtaining structured JSON data.

[0139] The following further elaborates on the establishment of structured JSON data according to the basic data information and the domain rule library of the present application by way of example. It should be understood that this example does not constitute any limitation to the present application.

[0140] Prompt word design:

[0141] Role injection: Input prompt words into the large model, describing its role as "expert in building a hydroformylation system diagram database, proficient in catalyst design, reaction kinetics and process optimization";

[0142] Rule embedding: Convert the domain rule base into constraints described in natural language (e.g., "When encountering the expression 'selectivity > 100%', it must be marked as abnormal data");

[0143] Through the confidence scoring mechanism, the model can learn the judgment tendencies of domain experts. For example, when cutting-edge terms such as "quantum size effect" are detected, the strict verification threshold is lowered;

[0144] Format instruction: The model output must contain the "Entity Type_Source Document ID_Confidence" field, for example: Entity Type: RhCl(PPh3)3, Type: Catalyst, Attributes: Particle size 5nm, Support Activated Carbon, Source: DOI:10.1021 / acs.catal.5b00156, Confidence: 0.95;

[0145] Task decomposition and execution:

[0146] Input the text data to be processed into the large model to obtain a set of triples, for example:

[0147] Input: Original text paragraph (e.g., "RhCl3 was loaded onto an Al2O3 carrier by an impregnation method and calcined at 400°C to obtain a catalyst")

[0148] Processing: The large model identifies "RhCl3" as the precursor, "Al2O3" as the carrier, "400°C" as the process parameter, and associates it with the "impregnation method" preparation method;

[0149] Output: a set of triplets (RhCl3 is the precursor, Al2O3 is the support, and 400°C is the calcination temperature).

[0150] Process analysis:

[0151] Input: Description of the experimental steps (e.g., "immerse the support in Rh solution for 2 hours, dry at 110°C, and then calcine in air at 500°C")

[0152] Processing: The large model was disassembled into the process chain of "impregnation (2h) → drying (110℃) → calcination (500℃)" and the temperature sequence was verified to be in compliance with the rule of "drying temperature is lower than calcination temperature";

[0153] Output: Process sequence with timestamps (impregnation step, drying step, calcination step, satisfying temperature sequence constraints).

[0154] Performance indicator association:

[0155] Input: Performance data segment (e.g. "Catalyst A has a TOF of 1200 h at 80 °C" -1 , selectivity 92%”)

[0156] Processing: Extract numerical data from large model and associate units (TOF = 1200h -1 ), and at the same time compare the “reasonable selectivity range (80% to 99%)” in the rule base;

[0157] Output: Structured performance record (TOF value 1200, unit h -1 , reaction temperature 80°C, selectivity 92%, data validity compliance);

[0158] Dynamic rule verification (input: preliminary extraction results; output: verification mark data).

[0159] When "catalyst particle size 0.5nm" is detected, the model automatically triggers the "nanomaterial particle size lower limit 1nm" constraint in the rule library and marks it as "suspected quantum effect description."

[0160] When encountering "reaction temperature -20℃", a special tag of "low-temperature process innovation" is generated based on common knowledge in the field (hydroformylation reaction requires high temperature), retaining the original data but reducing the confidence level.

[0161] Through the above processing, the information output by the large model is collected to form a structured data set.

[0162] In this embodiment, the structured JSON data (including entities, relationships, attributes, and document sources) may include the following:

[0163] Basic layer: entity-relationship-attribute triples (e.g., “Catalyst A is prepared using the impregnation method”);

[0164] Rule layer: compliance annotation (e.g., “temperature data complies with process logic constraints”);

[0165] Traceability layer: original text fragment mapping (such as "400℃ roasting" is associated with the second line of the third paragraph of the document).

[0166] In this embodiment, the large model can also output the following content:

[0167] Semantic conflict report, which may include the following:

[0168] Entity conflict: "Document X states that the particle size of catalyst B is 5nm, while Document Y states it is 500nm";

[0169] Numerical contradiction: "The chart shows 85% selectivity, but the text describes it as 95%"

[0170] Process paradox: "Roasting first, then impregnation" conflicts with the industry rule that "impregnation must occur before roasting".

[0171] In this embodiment, the large model can also output the following content:

[0172] Dynamic adaptation records (model learning logs), for example:

[0173] New term: "Quantum Size Catalyst" (automatically creates a new node type when particle size < 1nm is detected);

[0174] Rule adjustment: "Change the judgment threshold of selectivity > 100% from strict rejection to warning label"

[0175] Confidence calibration: "Data from the journal J. Catal." has a default confidence increase of 15%."

[0176] In this embodiment, large models suitable for this application include GPT-3.5, GPT-4, BERT, T5, etc. The selection criteria include the model's language understanding ability, domain adaptability, and computing resource requirements.

[0177] In this embodiment, the steps of constructing a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data, include:

[0178] Performing an integrity check on the structured JSON data, thereby obtaining an integrity check report and updated JSON data;

[0179] Perform domain rule filtering based on the updated JSON data and the domain rule library to obtain filtered data and rule triggering records;

[0180] Performing cross-validation enhancement based on the filtered data and the domain rule library to obtain cross-validation results and enhanced data;

[0181] The domain rule base is updated according to the enhanced data, thereby obtaining an updated domain rule base.

[0182] The following further illustrates the structured data quality verification and dynamic enhancement of the present application in detail by way of examples. It should be understood that the examples do not constitute any limitation to the present application.

[0183] Load entity types and required attribute mapping tables from the knowledge graph infrastructure document;

[0184] Use the JSON Schema validation tool to traverse each JSON data, record missing fields, and add a quality label to each JSON data, such as {"quality label":"incomplete"}, to generate a completeness check report and updated JSON data.

[0185] The performing domain rule filtering according to the updated JSON data and the domain rule library to obtain filtered data and rule triggering records includes:

[0186] Convert the rule base into executable logic (e.g. convert "mark as anomaly when temperature > 1000°C" into iftemperature>1000thenmark_as_anomaly());

[0187] Numeric range validation: Apply range rules to numeric attributes and flag abnormal data (e.g., {"temperature":1200,"unit":"℃","abnormal flag":"R001"});

[0188] Process logic verification: Use finite state machine (FSM) to verify the process sequence and mark logic conflicts (e.g. "baking first, then impregnation" is marked as "logic conflict");

[0189] Conflict resolution: If multiple rules conflict, the decision is made based on the priority of the rules;

[0190] Through the above processing, the filtered data and rule triggering records are obtained.

[0191] In this embodiment, performing cross-validation enhancement based on the filtered data and the domain rule library to obtain cross-validation results and enhanced data includes:

[0192] Verify consistency within the document: Extract causal relationships in the text (e.g., "High temperature leads to reduced selectivity"), compare text descriptions with chart data (e.g., text says "selectivity 80%", chart shows "60%"), and use NLP similarity algorithms to calculate semantic consistency;

[0193] Cross-document entity alignment: String matching + context features (such as preparation methods, reaction conditions) are used to align entities with the same name, with high-resolution characterization data being preferred (e.g., TEM data over SEM data).

[0194] Dynamic weight adjustment: Calculate data credibility weight based on literature sources;

[0195] In this embodiment, the cross-validation result is in JSON format, including consistency score, alignment decision, and weight value, for example, {"consistency score": 0.5, "alignment decision": "adopt document 1 data"}).

[0196] The enhanced data is in JSON format and contains weight annotations, such as "data credibility": 0.85}.

[0197] In this embodiment, updating the domain rule base according to the enhanced data, thereby obtaining an updated domain rule base includes:

[0198] Experts make decisions on each contradictory data pair to form new rules, and then add the new rules to the domain rule base or delete the previous rules, thereby updating the domain rule base.

[0199] In this embodiment, mapping and preprocessing the verified structured data from the structured data to the graph to obtain preprocessed graph data includes:

[0200] Obtaining a graph mapping rule table and preliminary mapped graph data based on the verified structured data;

[0201] Generate standardized atlas data and preprocessing logs based on the initially mapped atlas data;

[0202] Conflict detection and resolution are performed based on the standardized atlas data to obtain preprocessed atlas data.

[0203] The mapping and preprocessing of the verified structured data to the graph in the present application to obtain the preprocessed graph data is further elaborated below by way of example. It is to be understood that this example does not constitute any limitation to the present application.

[0204] In this embodiment, obtaining a graph mapping rule table and preliminary mapped graph data based on the verified structured data includes:

[0205] Mapping rule definition:

[0206] According to the graph framework documentation, align the fields in the structured data with the graph elements. For example, map the "catalyst name" field in the JSON to the graph node label and the "preparation method" field to the edge type.

[0207] Define many-to-one mapping logic (e.g., "nanoparticle" and "nanocrystal" are both normalized to the "nanostructure" node type);

[0208] Relationship type mapping:

[0209] Convert the implicit logical relationships in the text into explicit graph edge types (such as "drive" and "solve"). For example, map "high temperature promotes reaction rate" to the edge of "high temperature → drive → reaction rate improvement".

[0210] Attribute standardization:

[0211] Add unit constraints for numeric attributes (e.g. "activity data must be marked with h -1 or mol / g units”), and create independent nodes for non-numeric attributes (e.g., “selectivity > 90%” is split into a “selectivity” node and a “value range” relationship).

[0212] Through the above steps, we can obtain the graph mapping rule table (the mapping relationship between record fields and graph elements) and the preliminary mapped graph data (including the temporary structure of nodes, edges, and attributes);

[0213] In this embodiment, generating standardized atlas data and preprocessing logs based on the initially mapped atlas data includes:

[0214] Data standardization:

[0215] Normalize the units in the preliminary mapped atlas data. For example, convert "5nm" and "5000 microns" to "5e-9 meters" and mark the original units.

[0216] Normalize the terms in the preliminary mapped atlas data, for example, merge synonyms (e.g., "Pt / C catalyst" and "platinum carbon catalyst" point to the same node);

[0217] Process missing values ​​in the initially mapped atlas data, for example, fill in placeholders for missing attributes (e.g., mark the missing “particle size” attribute as “unknown”);

[0218] For data with missing relationship types in the preliminary mapped graph data, the default relationship is inferred based on the context (e.g., “catalyst A → unknown relationship → reaction B” is supplemented with a “participation” relationship).

[0219] Clean the format of the initially mapped atlas data, for example, remove redundant spaces and special characters (such as "Catalyst A (new)" is cleaned to "Catalyst A"), and unify the date format (such as "2025.6.9" and "2025-06-09" are converted to standard ISO format).

[0220] Through the above processing, standardized atlas data and preprocessing logs can be obtained.

[0221] In this embodiment, conflict detection and resolution are performed based on the standardized atlas data to obtain pre-processed atlas data, including:

[0222] Conflict Detection:

[0223] Detect possible inconsistencies in normalized atlas data, such as:

[0224] The attributes of the same entity in different documents are inconsistent (e.g., document 1 states that “the particle size of catalyst A is 5 nm”, while document 2 states that it is “500 nm”);

[0225] Detection of logical conflicts in relationships (e.g., the coexistence of "Method A → Suppress → Activity" and "Method A → Enhance → Activity");

[0226] Conflict Resolution:

[0227] Automatic decision-making: judge based on the priority of the rule base (e.g., retaining experimental data when “experimental data > literature conclusions”);

[0228] Manual intervention: For conflicts that cannot be resolved by high-priority rules, a visual comparison interface is generated (e.g., displaying the experimental conditions of two papers side by side);

[0229] Recording and feedback:

[0230] Mark the conflict resolution results (e.g., "The data from Reference 1 was adopted because its experimental conditions were more stringent");

[0231] Updated the rule base (e.g., added "When the document source is Nature Catalysis, the data priority is +30%").

[0232] Through the above processing, the preprocessed atlas data is obtained.

[0233] In this embodiment, the multimodal data fusion and association of the pre-processed atlas data and the experimental chart data to obtain a preliminary multimodal fusion atlas includes:

[0234] Processing the experimental chart data to obtain structured chart data;

[0235] Performing cross-modal association on the structured chart data and the pre-processed atlas data, thereby obtaining cross-modal association data;

[0236] A modality fusion atlas is obtained according to the cross-modal association data and the preprocessed atlas data.

[0237] The following further illustrates the method of obtaining a preliminary multimodal fusion map in the present application by way of example. It should be understood that this example does not constitute any limitation to the present application.

[0238] In this embodiment, the experimental chart data can be image data or tabular data in various documents, for example, structured charts such as TEM images (such as catalyst particle size distribution diagrams) and XRD spectra (such as interplanar spacing analysis diagrams) in literature related to hydroformylation reactions, and the charts contain numerical data (such as particle size ranges, interplanar spacing values) and non-numerical annotations (such as legends and coordinate axis labels).

[0239] In this embodiment, the following method may be used to process the experimental chart data to obtain structured chart data:

[0240] Chart detection and positioning:

[0241] Use deep learning-based object detection models (such as Faster R-CNN) to identify the bounding boxes of charts in documents and separate them from the text content;

[0242] For hand-drawn sketches or fuzzy charts, a cascade detection strategy is adopted: first, a lightweight model (such as MobileNet) is used to quickly screen candidate areas, and then a high-precision model (such as ResNet-50) is used for secondary verification.

[0243] Chart Type Classification:

[0244] Use convolutional neural networks (CNNs) to identify chart types (e.g., line charts, bar charts, scatter plots). For example, for a TEM image, classify it as a "particle size distribution chart" to trigger the particle size extraction logic; for an XRD spectrum, classify it as a "crystal spacing chart" to trigger the peak position detection logic.

[0245] Data point extraction:

[0246] Numerical data: For TEM images, the CNN-based method proposed by Poco et al. is used to locate the boundaries of catalyst particles, and the pixel positions are converted into actual particle size values ​​(e.g., 5 nm) in combination with a linear interpolation algorithm. Non-numeric annotation: OCR technology is used to extract legends and coordinate axis labels (e.g., "particle size / nm") and associate them with numerical data.

[0247] Legend - Data Series Binding:

[0248] Use color similarity matching algorithms (such as the CIEDE2000 color difference formula) to associate legends with data series to solve the problem of correspondence between text and visual elements. For example, bind the "Catalyst A" legend in the TEM image to the data series of particle size 5nm.

[0249] Through the above processing, structured chart data can be obtained.

[0250] In this embodiment, performing cross-modal association on the structured chart data and the pre-processed atlas data to obtain cross-modal association data includes:

[0251] Features of the structured graph data and the preprocessed graph data are extracted separately. For example, BERT is used to extract semantic features of entities in the preprocessed graph (such as catalyst names), and CNN is used to extract visual features of the graph data (such as particle size distribution histograms).

[0252] Perform feature alignment on the extracted features, for example, by mapping text features and chart features into a unified space through common subspace learning (such as canonical correlation analysis, CCA), and calculating cosine similarity, for example, aligning the text features of "catalyst A" with the visual features of the particle size of 5nm in the TEM image.

[0253] Perform entity association. For example, associate the numerical value extracted from the chart (such as particle size 5nm) with the catalyst node attribute in the preprocessed map, establish the relationship of "catalyst A→particle size 5nm", and for entities with the same name (such as "catalyst A" in different documents), determine the final attribute value through majority voting or expert judgment.

[0254] Through the above processing, cross-modal related data is obtained.

[0255] In this embodiment, obtaining a modality fusion atlas according to the cross-modal association data and the pre-processed atlas data includes:

[0256] Integrate entities, relationships, and attributes in cross-modal relational data into the preprocessed graph. For example, integrate "Catalyst A → Particle size 5nm (source: TEM image)" into the catalyst node attributes of the graph.

[0257] By integrating graph data, the attribute information of entities is enhanced. For example, the catalyst node contains not only textual attributes (such as name and type), but also numerical attributes extracted from the graph (such as particle size and interplanar spacing).

[0258] Update the integrated data into the graph to form a complete knowledge graph containing multimodal information.

[0259] In this embodiment, the multimodal fusion graph (including experimental graph metadata) includes: entities, relationships, attributes and source metadata.

[0260] In this embodiment, generating an enhanced knowledge graph based on the multimodal fusion preliminary graph includes:

[0261] Using the breadth-first search algorithm, we randomly select an entity node from the preliminary multimodal fusion graph as the starting point, traverse its directly related nodes and indirectly related nodes layer by layer, and record all entity-relationship combinations on the traversal path;

[0262] Count the number of occurrences of entity-relationship combinations in all traversal paths, and obtain a list of high-frequency association patterns for each entity-relationship combination whose occurrence number exceeds a preset threshold;

[0263] Perform causal inference on the list of high-frequency association patterns to obtain a list of causal relationship candidates;

[0264] The multimodal fusion preliminary map is updated according to the causal relationship candidate list and the multimodal fusion preliminary map, thereby forming an enhanced knowledge map.

[0265] Specifically, using the breadth-first search algorithm, we randomly select an entity node from the preliminary multimodal fusion graph as the starting point, traverse its directly related nodes and indirectly related nodes layer by layer, and record all entity-relationship combinations on the traversal path, including:

[0266] Using the breadth-first search (BFS) algorithm, an entity node (such as "catalyst A") is randomly selected from the graph as the starting point, and its directly related nodes (such as "preparation method B" and "performance index C") and indirectly related nodes (such as "reaction conditions D") are traversed layer by layer, and all entity-relationship combinations on the traversal path are recorded (such as "catalyst A→preparation method B→particle size 5nm").

[0267] Count the number of occurrences of entity-relationship combinations in all traversal paths. For example, if “immersion method → ​​small particle size” appears 90 times in 100 documents, set a threshold (e.g., co-occurrence frequency > 80%) and mark the significant association pattern.

[0268] Match high-frequency association patterns with the inference rule base. For example, if the rule "particle size < 10nm → increased activity" is matched and the current mode is "immersion method → ​​small particle size (5nm)", record the successfully matched rule and corresponding mode.

[0269] Calculate the conditional probability of the matching pattern. For example, among the 90 documents matching "impregnation method → ​​small particle size", 80 also mentioned "activity improvement". The conditional probability is 80 / 90≈89%. Combined with the Bayesian network, the confidence level of the causal relationship is evaluated (such as P(activity improvement | impregnation method) = 89%).

[0270] Through the above processing, a list of causal relationship candidates is obtained (such as "immersion method → ​​causal relationship → activity improvement"), and each relationship is accompanied by a rule matching path, conditional probability and confidence score.

[0271] Check whether the candidate causal relationship is consistent with the existing data in the map. For example, if the map contains the path "immersion method → ​​particle size 5nm → activity increased by 100%", the verification is passed.

[0272] Generate an exception mark for contradictory results (such as "a certain document claims that the impregnation method causes the particle size to increase and the activity to decrease") and record the reasoning path for reference.

[0273] Convert the verified causal relationship into a new triple (e.g., “immersion method → ​​causal relationship → activity improvement”), and attach a confidence score and a list of supporting literature to the new triple.

[0274] See also Figure 3 as well as Figure 4 In the second embodiment, a method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction includes:

[0275] Step 1) Determine the basic framework of the knowledge graph by collecting professional corpus covering substance names, synthesis processes, performance tests, and application descriptions.

[0276] Step 2) Based on the predefined graph structure and domain terminology, a prompt instruction template was designed to guide the model in identifying key entities and relationships within the text. Leveraging the large language model's deep understanding of semantic context, the system automatically outputs structured graph units such as entity attributes and relationship triples, encompassing the core knowledge logic of the catalytic system.

[0277] Step 3) Use a script to import the structured information generated by the large language model into the Neo4j database to complete the construction of a knowledge graph.

[0278] The detailed process of step 1) is as follows:

[0279] Before constructing the knowledge graph of any professional system, it is necessary to first systematically collect the core literature in the field and conduct a structured analysis of its content in order to extract a basic framework of the knowledge graph suitable for automatic extraction. The framework logic proposed in the present invention simulates the cognitive path of scientific researchers, especially graduate students, in the process of literature reading: that is, starting from grasping the research background and the key problems to be solved, and gradually focusing on the key information such as the technical innovations, the material systems used, the experimental methods, and the performance reflected in the literature. The above content not only constitutes the semantic basis for the design of prompt instructions and the construction of the input context, but also provides a stable and efficient structural guidance for the subsequent knowledge extraction module. With the help of the graph framework designed by this logic, it can be closer to the information acquisition habits and decision-making processes of scientific researchers in the field of chemical engineering and materials in actual work, thereby significantly improving the practicality and interpretability of the knowledge graph.

[0280] The detailed process of step 2) is as follows:

[0281] Based on the framework of step 1), we design knowledge graph extraction prompt instructions, define entity and relationship extraction rules, and standardize the output format. We provide different examples based on different needs, prompting the large model to automatically extract entities and relationships in the text, and any standardized output triples.

[0282] 2.1) Providing system prompts to let the model act as a domain expert, so that it can acquire rich domain knowledge;

[0283] 2.2) Based on the framework determined in step 1), a specific task description is made to guide the large model to extract entities and relationships;

[0284] 2.3) Design an extraction format to output the entities and relations extracted from the text in the form of triples;

[0285] 2.4) Provide different data examples so that the large model can understand the output styles for different aspects;

[0286] 2.5) Integrate the task description, extraction format, and data examples to form a complete knowledge extraction instruction. The instruction will serve as an important reminder for text extraction, ensuring that the content extracted by the large model meets the requirements;

[0287] 2.6) Combine the above instructions with the text to be extracted and provide them as input to the large model;

[0288] 2.7) The large model extracts entities and relationships based on the prompt instructions and input text content, and outputs them in triple format according to the requirements of the extraction format.

[0289] The detailed step 3) includes the following specific steps:

[0290] 3.1) Merge entity nodes with the same name and add document source information;

[0291] 3.2) Use Cypher statements to import the standardized structured data into the Neo4j database.

[0292] Compared with traditional knowledge graph construction methods that rely on manual annotation or rule-based matching, the present invention realizes the automation of the entire process from document collection, semantic extraction to graph storage through the deep integration of large language models and prompt engineering. It not only significantly reduces labor costs and knowledge engineering thresholds, but also introduces an entity attribute extension mechanism based on standard triples, so that quantitative factors such as material properties and process parameters can be simultaneously incorporated into the graph structure, significantly enhancing the richness and professional depth of knowledge expression to meet the needs of professional personnel.

[0293] In addition, the prompt template system and scripted processing flow designed by this method are highly versatile and can be flexibly adapted to multiple sub-fields such as inorganic materials, polymers, and battery materials. They support rapid migration without the need to reconstruct the underlying system architecture. The automatic binding mechanism of document metadata is introduced in the entity node processing link, which effectively improves the verifiability and update controllability of the graph. The unified template management mechanism and cleaning and import process ensure the robustness and continuous expansion capabilities of the system in response to terminology evolution and ontology changes. In summary, while improving the level of intelligent graph construction, the present invention provides a new technical path and system framework for knowledge organization, structural expression, and causal reasoning in the field of chemical engineering and materials.

[0294] like Figure 2 As shown, a knowledge graph construction method based on a large model prompting project includes the following steps:

[0295] (1) Determination of the knowledge graph framework

[0296] After an in-depth study of multiple papers related to hydroformylation systems, we determined that the knowledge graph needed to cover elements such as catalyst composition, preparation process, characterization results, performance testing, as well as key issues and innovations addressed, and used this as the basis for constructing prompt words.

[0297] (2) Design of prompt words for hydroformylation system and output of triples

[0298] 2.1) Based on the determined knowledge graph framework, design a special prompt as follows:

[0299] You are an expert in building graph databases for hydroformylation systems. You are proficient in extracting knowledge from technical text to construct graph-based structured data. You are particularly skilled in identifying entities, constructing nodes, annotating attributes, and establishing relationships between entities. Please strictly follow the following requirements and output a valid, reproducible JSON structure (JSON only, without any explanations or comments).

[0300] Please complete the following tasks based on the text provided:

[0301] Task 1: Named Entity Recognition

[0302] 1. Extract all entities with knowledge graph value from the text (such as materials, compounds, intermediates, species, catalysts, additives, carriers, characterization methods, performance indicators, etc.).

[0303] Task 2: Entity Classification

[0304] 1. Assign a Wikidata-style class to each entity (e.g., Catalyst, Material, Condition, etc.) to reflect its ontological meaning.

[0305] Task 3: Attribute annotation specifications

[0306] 1. "properties" only stores attribute values ​​of digital type (such as value, range, percentage, concentration, temperature, etc.).

[0307] 2. All non-numeric text descriptions (such as color, appearance, purpose, concluding remarks, etc.) must be extracted as independent entity nodes and associated with the corresponding entity through appropriate relationships (class="Attribute" or more specific categories can be used).

[0308] 3. If the entity lacks any numeric properties, the "properties" field should be an empty object "{}".

[0309] 4. All attribute names must use camelCase naming format to express clear feature meanings.

[0310] 5. If there are multiple values ​​for the same attribute, use array format to represent them uniformly.

[0311] 6. It is prohibited to fabricate or speculate on the attributes of entities. Please leave missing items blank.

[0312] 7. If the extracted numbers have units, be sure to include them.

[0313] Task 4: Preparation process refinement

[0314] 1. If the text contains preparation methods (such as "immersion method" or "hydrothermal method"), please generate an entity node for each method, with class="PreparationMethod" and abbreviation="None" (if there is no abbreviation).

[0315] 2. For the specific process of the method in the same literature, extract its key parameters and units, and strictly indicate what substances were added, including but not limited to:

[0316] -Solvent (e.g. THF, 50 mL)

[0317] -Precursor (e.g. Rh(acac)(CO)2, 0.2g)

[0318] - Additives, intermediates, reducing agents, etc.

[0319] 3. Establish the following relationships:

[0320] - Preparation method → ​​"catalyst" preparation process, "relation": "preparation process"

[0321] - "catalyst" → "catalyst" preparation process, "relation": "prepared from"

[0322] - "catalyst" is replaced by the catalyst in the specific article

[0323] for example:

[0324]

[0325]

[0326] ],

[0327] "calcinationTemperature":[

[0328] "500℃"

[0329] ],

[0330] "calcinationTime":[

[0331] "2h"

[0332] ],

[0333] "reductionTemperature":[

[0334] "500℃"

[0335] ],

[0336] "reductionGas":[

[0337] "Ar / H2(10 / 10ml min-1)" ]

[0339] }

[0340] },

[0341] "to":{

[0342] "entity":"Incipient Wetness Impregnation",

[0343] "abbreviation":"None",

[0344] "class":"PreparationMethod",

[0345] "properties":{}

[0346] }

[0347] }

[0348] 4. There is no need to establish an additional reverse relationship from PreparationProcess to PreparationMethod, nor is there a need for Catalyst to directly connect to PreparationMethod.

[0349] 5. The preparation process only generates one node, so the internal information of this node must be complete

[0350] 6. The added precursors must include not only the amount but also what kind of precursors were added.

[0351] Task 5: Industry Needs, Research Challenges, and Technological Innovation

[0352] 1. Including but not limited to extracting and constructing the following entities:

[0353] -IndustryNeed (Industry Need)

[0354] -Challenge

[0355] -Innovation

[0356] 2. Including but not limited to extracting and building the following relationships:

[0357] -IndustryNeed→Challenge, "relation": "Drive"

[0358] -Challenge→Innovation, "relation": "solve"

[0359] -Innovation→PreparationMethod / Material, etc., "relation": "propose a solution"

[0360] -If not, do not extract.

[0361] -If yes, please include how this article achieves these, such as: what is achieved through preparation method, what is achieved through what material, what is achieved through what technology.

[0362] Task 6: Performance Metrics

[0363] 1. Performance indicators (such as selectivity, conversion, yield, TOF, lifetime, etc.) should be classified as "Performance" category rather than directly attributed to the catalyst or reaction condition node.

[0364] - In the "Performance" category node, use properties to list performance indicators, and put values ​​or ranges into arrays.

[0365] 2. If the text describes a conclusion drawn from the catalytic performance:

[0366] -Starting from the catalytic performance node, establish the corresponding cause and effect relationship.

[0367] Task 7: Representation methods and representation results

[0368] 1. Generate a node for each characterization method, class="CharacterizationMethod" (such as FTIR, TEM, XRD).

[0369] 2. Directly establish the relationship between catalyst and characterization method, "relation": "characterization", indicating that a certain characterization method has been applied to the catalyst.

[0370] 3. The characterization results should be recorded in the "properties" of the characterization method node (such as particle size, peak position, morphology, etc.), but the characterization method itself is not used as a value in the "properties".

[0371] 4. If the text describes an inference drawn from the representation results:

[0372] -Starting from the characterization method node, establish the causal relationship between performance changes, structural changes, etc.

[0373] The output JSON format is as follows (only JSON is output without any explanatory text): [

[0375]

[0376] Additional constraints:

[0377] 1. Entity naming requirements are concise: Node names should be limited to 1-3 words and use academic or general terms (such as Rh(CO)2, Co / SiO2, H2 / CO).

[0378] 2. If the entity has no abbreviation, the "abbreviation" field is filled with "None".

[0379] 3. If there are parent and derivatives (such as CO and Rh(CO)2) in the same system, create nodes for each.

[0380] 4. All relationships must actually exist in the text and must not be inferred or fabricated.

[0381] 5. Cover the following knowledge dimensions (if mentioned in the text):

[0382] -Preparation process

[0383] -Active ingredient

[0384] - Additives / Carriers

[0385] -Reaction conditions

[0386] -Catalytic performance

[0387] -Representational methods

[0388] -Physical characterization results

[0389] -Causal inference between performance and structure

[0390] 6. The output results should be reproducible and stable to ensure that the output structure is consistent under the same input.

[0391] 7. Knowledge graph goal: Assist in designing reasonable preparation formulas based on catalytic performance requirements.

[0392] 8. Extract as much data from the article as possible.

[0393] 9. The catalyst itself does not have any "properties" and exists only as a node.

[0394] 10. You can establish multi-layer causal relationships and find inferential statements.

[0395] 11. Please think carefully about whether your answer is correct before outputting it to ensure its accuracy.

[0396] 12. The preparation process must be described in detail, the extracted information must be reproducible, and the numbers with units must include the units.

[0397] 13. The selected preparation method must be a catalyst synthesis method, such as impregnation method.

[0398] 14. All output must be in English, and Chinese is not allowed.

[0399] 2.2) Upload the prompt words and documents to the deepseek-r1 model and let it extract triples according to the requirements.

[0400] (3) Importing knowledge graph

[0401] 3.1) Merge the "properties" of entities with the same name found in the same document to ensure the consistency of information;

[0402] 3.2) Insert "DOI" and "Literature" attributes for each entity to ensure that each node has a source;

[0403] 3.3) Import the processed structured data into entities and relationships through cypher statements to complete the construction of the knowledge graph.

[0404] Compared with traditional knowledge graph construction methods that rely on manual annotation or rule-based matching, the present invention realizes the automation of the entire process from document collection, semantic extraction to graph storage through the deep integration of large language models and prompt engineering. It not only significantly reduces labor costs and knowledge engineering thresholds, but also introduces an entity attribute extension mechanism based on standard triples, so that quantitative factors such as material properties and process parameters can be simultaneously incorporated into the graph structure, significantly enhancing the richness and professional depth of knowledge expression to meet the needs of professional personnel.

[0405] In addition, the prompt template system and scripted processing flow designed by this method are highly versatile and can be flexibly adapted to multiple sub-fields such as inorganic materials, polymers, and battery materials. They support rapid migration without the need to reconstruct the underlying system architecture. The automatic binding mechanism of document metadata is introduced in the entity node processing link, which effectively improves the verifiability and update controllability of the graph. The unified template management mechanism and cleaning and import process ensure the robustness and continuous expansion capabilities of the system in response to terminology evolution and ontology changes. In summary, while improving the level of intelligent graph construction, the present invention provides a new technical path and system framework for knowledge organization, structural expression, and causal reasoning in the field of chemical engineering and materials.

[0406] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the present invention. Therefore, such modifications and improvements, which do not depart from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction, characterized in that: The method for constructing a knowledge graph in the field of chemical materials based on the inference large model prompt instruction includes: Obtaining basic data information and experimental chart data, wherein the basic data information includes professional literature in the chemical industry and a field terminology system; Establish a domain rule base and a knowledge graph basic framework document based on the basic data information; Creating structured JSON data based on the basic data information and the domain rule library, wherein the structured JSON data includes entities, relationships, and attributes; Building a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data; Mapping and preprocessing the verified structured data to graphs to obtain preprocessed graph data; Perform multimodal data fusion and association on the preprocessed atlas data and experimental chart data to obtain a preliminary multimodal fusion atlas; Establish domain knowledge reasoning association data for the multimodal fusion preliminary graph, thereby obtaining a knowledge graph containing reasoning results.

2. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 1, characterized in that: The establishment of a domain rule base and a knowledge graph basic framework document based on the basic data information includes: Obtain original document collection; Perform text preprocessing on the original document collection to obtain text data that can be used for recognition; Extract features from text data that can be used for recognition to obtain core feature data; Performing a logical chain construction on the core feature data to obtain logical chain data; wherein the logical chain data and the core feature data constitute structured document data; Define node types, relationship types, and attribute rules for structured document data to obtain knowledge graph framework information; Generate rule template library based on domain terminology system; A rule priority table is generated according to the rule template library; the rule template library and the rule priority table constitute a domain rule library.

3. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 2, characterized in that: The step of establishing structured JSON data based on the basic data information and the domain rule library includes: Prompt word design, rule embedding design, format instruction design, and entity type design are performed according to the basic data information and the domain rule library, thereby generating a structured instruction template; Decomposing the task according to the basic data information and the structured instruction template to obtain structured semantic units; The structured semantic unit is converted into JSON format, thereby obtaining structured JSON data.

4. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 3, characterized in that: The step of constructing a Neo4j database based on the structured JSON data and the domain rule base to perform structured data quality verification and dynamic enhancement, thereby obtaining an updated domain rule base and verified structured data, includes: Performing an integrity check on the structured JSON data, thereby obtaining an integrity check report and updated JSON data; Perform domain rule filtering based on the updated JSON data and the domain rule library to obtain filtered data and rule triggering records; Performing cross-validation enhancement based on the filtered data and the domain rule library to obtain cross-validation results and enhanced data; The domain rule base is updated according to the enhanced data, thereby obtaining an updated domain rule base.

5. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 4, characterized in that: The mapping and preprocessing of the verified structured data to a graph to obtain preprocessed graph data includes: Obtaining a graph mapping rule table and preliminary mapped graph data based on the verified structured data; Generate standardized atlas data and preprocessing logs based on the initially mapped atlas data; Conflict detection and resolution are performed based on the standardized atlas data to obtain preprocessed atlas data.

6. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 5, characterized in that: The multimodal data fusion and association of the pre-processed atlas data and the experimental chart data to obtain a preliminary multimodal fusion atlas includes: Processing the experimental chart data to obtain structured chart data; Performing cross-modal association on the structured chart data and the pre-processed atlas data, thereby obtaining cross-modal association data; A modality fusion atlas is obtained according to the cross-modal association data and the preprocessed atlas data.

7. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 6, characterized in that: The step of establishing domain knowledge reasoning association data for the multimodal fusion preliminary graph to obtain a knowledge graph containing reasoning results includes: generating an enhanced knowledge graph based on the multimodal fusion preliminary graph; Construct a knowledge graph containing reasoning results based on the enhanced knowledge graph.

8. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 7, characterized in that: Generating an enhanced knowledge graph based on the multimodal fusion preliminary graph includes: Using the breadth-first search algorithm, we randomly select an entity node from the preliminary multimodal fusion graph as the starting point, traverse its directly related nodes and indirectly related nodes layer by layer, and record all entity-relationship combinations on the traversal path; Count the number of occurrences of entity-relationship combinations in all traversal paths, and obtain a list of high-frequency association patterns for each entity-relationship combination whose occurrence number exceeds a preset threshold; Perform causal inference on the list of high-frequency association patterns to obtain a list of causal relationship candidates; The multimodal fusion preliminary graph is updated according to the causal relationship candidate list and the multimodal fusion preliminary graph to form an enhanced knowledge graph, wherein the enhanced knowledge graph includes the multimodal fusion preliminary graph and new causal relationship edges generated by the causal relationship candidate list.

9. The method for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction according to claim 8, characterized in that: The knowledge graph containing the reasoning results constructed based on the enhanced knowledge graph includes: Import the enhanced knowledge graph into the graph database, preserving the original document source and reasoning path metadata; Merge entity nodes with the same name to generate a knowledge graph containing reasoning results.

10. A device for constructing a knowledge graph in the field of chemical materials based on a large inference model prompt instruction, characterized in that: The device for constructing a knowledge graph in the field of chemical materials based on the inference large model prompt instruction includes: A basic data acquisition module, which is used to acquire basic data information and experimental chart data, wherein the basic data information includes professional literature in the chemical industry and a field terminology system; An establishment module, wherein the establishment module is used to establish a domain rule library and a knowledge graph basic framework document based on the basic data information; A structured JSON data acquisition module, wherein the structured JSON data acquisition module is used to establish structured JSON data according to the basic data information and the domain rule library, wherein the structured JSON data includes entities, relationships and attributes; A quality verification module, which is used to construct a Neo4j database based on the structured JSON data and the domain rule library to perform quality verification and dynamic enhancement of the structured data, thereby obtaining an updated domain rule library and verified structured data; A mapping module, which is used to map and preprocess the verified structured data from the structured data to the graph, thereby obtaining preprocessed graph data; An association module is used to perform multimodal data fusion and association on the pre-processed atlas data and experimental chart data, thereby obtaining a preliminary multimodal fusion atlas; The reasoning association module is used to establish domain knowledge reasoning association data for the multimodal fusion preliminary graph, thereby obtaining a knowledge graph containing reasoning results.

Images