Construction of ultra-high performance concrete knowledge graph and machine interpretation method based on graph retrieval enhanced generation

Abstract

The application discloses a method for constructing an ultra-high performance concrete knowledge graph and machine explanation based on graph retrieval enhanced generation, and belongs to the technical field of ultra-high performance concrete knowledge graph construction and mechanism cognition, and comprises the following steps: obtaining a plurality of literatures related to ultra-high performance concrete; constructing an ultra-high performance concrete knowledge graph based on the plurality of literatures related to ultra-high performance concrete; performing retrieval of an adaptive length path on a target start and end node based on the ultra-high performance concrete knowledge graph, and obtaining at least one first candidate evidence chain corresponding to the target start and end node; inputting a target instruction and each first candidate evidence chain into a generation model to obtain corresponding machine explanation text. The method can efficiently obtain machine explanation of ultra-high performance concrete knowledge with high credibility.

Description

Technical Field

[0001] This invention belongs to the field of UHPC knowledge graph construction and mechanism cognition technology, specifically involving a mechanism explanation method for constructing a knowledge graph of ultra-high performance concrete and based on graph retrieval enhancement generation. Background Technology

[0002] Ultra-High Performance Concrete (UHPC) is a high-strength, high-durability, and low-porosity cementitious material. Due to its superior mechanical properties and durability, it is increasingly widely used in civil engineering, construction, and defense. Currently, knowledge about UHPC is mainly scattered across a vast amount of academic literature, patent texts, and technical reports, exhibiting a typical characteristic of "rich data but isolated knowledge." Traditional database and keyword retrieval methods struggle to effectively mine and utilize the complex connections hidden within these texts, leading to a dilemma of low efficiency in knowledge acquisition for researchers and engineers. Furthermore, traditional target large language models suffer from problems such as illusions and factual errors when generating content in specialized fields. Therefore, how to efficiently obtain credible mechanistic explanations of UHPC knowledge is a pressing technical problem that needs to be solved.

[0003] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0005] This disclosure provides a method for constructing a knowledge graph of ultra-high performance concrete and for generating a mechanism explanation based on graph retrieval enhancement, so as to efficiently obtain a mechanism explanation of ultra-high performance concrete knowledge with high credibility.

[0006] In some embodiments, the method for constructing a knowledge graph of ultra-high performance concrete and explaining the mechanism based on graph retrieval enhancement includes: Obtain multiple documents related to ultra-high performance concrete; Based on multiple literatures related to ultra-high performance concrete, a knowledge graph of ultra-high performance concrete is constructed. Based on the knowledge graph of ultra-high performance concrete, an adaptive length path retrieval is performed on the target origin and destination nodes to obtain at least one first candidate evidence chain corresponding to the target origin and destination nodes; The target instruction and each first candidate evidence chain are input into the generative model to obtain the corresponding mechanism explanation text.

[0007] The beneficial effects of this invention are as follows: By acquiring multiple documents related to ultra-high performance concrete (UHVPC) and constructing a knowledge graph of UHVPC, the powerful semantic association and reasoning capabilities of knowledge graph technology can be leveraged to connect fragmented UHVPC knowledge. The UHVPC knowledge graph is used to perform adaptive length path retrieval on the target start and end nodes to obtain at least one first candidate evidence chain, thus achieving efficient acquisition of the evidence chain corresponding to UHVPC knowledge. The target instruction and each first candidate evidence chain are then input into a generative model to obtain the mechanism explanation text output by the model. Multiple first candidate evidence chains provide the generative model with accurate and logically connected context, ensuring that the final generated mechanism explanation text is supported by the knowledge graph path and has high credibility. In this way, based on the constructed UHVPC knowledge graph, the evidence chain corresponding to UHVPC knowledge can be efficiently obtained, and then input into the generative model to obtain the mechanism explanation text corresponding to the highly credible UHVPC knowledge, thereby achieving efficient acquisition of highly credible mechanism explanations of UHVPC knowledge.

[0008] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0009] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a flowchart of a method for constructing a knowledge graph of ultra-high performance concrete and explaining the mechanism of graph retrieval-enhanced generation provided by the present invention; Figure 2 This is a flowchart of a method for updating a knowledge base provided by the present invention. Detailed Implementation

[0010] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0011] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0012] Unless otherwise stated, the term "multiple" means two or more.

[0013] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0014] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0015] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0016] For knowledge organization and design decisions related to ultra-high performance concrete (UHPC), various technical approaches have emerged in industry and academia: (1) Manual mapping based on ontology / dictionary and rules: Extract key points of literature using domain ontology, terminology dictionary and regular expressions / rules, and manually enter them into knowledge bases such as Neo4j / OWL to adapt to a small number of high-value topics. This approach is controllable on a small scale, but it is costly, has poor scalability, heavy maintenance burden, and is difficult to cover rapidly growing literature and engineering data.

[0017] (2) General Information Extraction (IE) Pipeline: NER / RE (such as BERT-CRF, pointer network, table extraction) is used to identify entities and relationships such as materials, processes, and properties from papers and reports, and write them into a graph database. This approach is effective for general corpora, but in the context of ultra-high performance concrete, it often lacks key attributes such as directionality (promoting / inhibiting) and strength, applicable conditions (dosage, age, curing); and there are significant differences in cross-document descriptions, resulting in insufficient standardized mapping.

[0018] (3) Large Language Model-Driven Extraction and Automatic Cypher Generation: Using instructional prompts, LLM can directly output triples or Cypher statements, which significantly improves the speed of map building and lowers the engineering threshold. However, LLM has uncertainties and "illusions", and is prone to errors in relation direction, units / dimensions, and conditional boundaries; if quality assessment and human-machine verification are lacking, it is difficult to directly input into the database to support engineering decisions.

[0019] (4) Knowledge integration and concept alignment: By using thesaurus, embedding alignment and clustering to merge synonyms such as “compressive strength / cubic compressive strength” and “glass powder / waste glass powder”, we attempted to unify entities and relationships. Due to the polysemy and rich complex expressions of ultra-high performance concrete terminology, and the different granularities of different papers, entity fragmentation and inconsistency of relationships are still prominent.

[0020] (5) Graph-driven retrieval and question answering / interpretation: Using path retrieval, subgraph matching, or GraphRAG, graph evidence is converted into explanatory text for explaining "why it is effective" and "under what conditions it is effective". Existing solutions lack a unified standard for the granularity and weight of the evidence chain, and the applicable boundaries and confidence levels are not structured, making it difficult to directly apply to parameter ranges and matching ratio recommendations.

[0021] (6) Data-driven blending optimization and knowledge-guided design: The blending search is carried out using methods such as machine learning / Bayesian optimization / multi-agent collaboration. Some systems have tried to use knowledge graphs for prior constraints or candidate generation. However, most of them have weak interpretability and the knowledge-experiment-design closed loop is not perfect. In the face of composition fluctuations and batch differences brought about by solid waste resource utilization, the versioning update and rollback capabilities are insufficient. The compatibility and mapping with enterprise LIMS / ELN and ledger databases are also insufficient.

[0022] Therefore, it is extremely important to efficiently obtain reliable mechanistic explanations of ultra-high performance concrete and related design solutions in the field of ultra-high performance concrete.

[0023] Combination Figure 1 As shown, this disclosure provides a method for constructing a knowledge graph of ultra-high performance concrete and explaining the mechanism of graph retrieval-based enhanced generation, including: Step S101: Obtain multiple documents related to ultra-high performance concrete.

[0024] Preferably, literature related to ultra-high performance concrete can be collected from academic databases, engineering reports, patent data, etc.

[0025] Step S102: Construct a knowledge graph of ultra-high performance concrete based on multiple documents related to ultra-high performance concrete.

[0026] Preferably, the knowledge graph of ultra-high performance concrete is updated every third preset time interval to achieve version updates.

[0027] Step S103: Based on the knowledge graph of ultra-high performance concrete, perform adaptive length path retrieval on the target origin and destination nodes to obtain at least one first candidate evidence chain corresponding to the target origin and destination nodes.

[0028] The target origin and destination nodes include the target start node and the target end node. The target origin and destination nodes are determined in response to user input.

[0029] In some embodiments, the target start and end nodes are user-inputted target start and end nodes related to the preparation of ultra-high performance concrete. For example, if the user's requirement is "how to improve the toughness of ultra-high performance concrete by adding steel fibers," then the corresponding target start node is: steel fibers; and the target end node is: toughness or, more specifically, high toughness. The evidence chain retrieval task is to find all paths in the ultra-high performance concrete knowledge graph that connect "steel fibers" and "toughness." These paths may contain nodes and relationships such as [steel fibers] - [bridging] -> [microcracks] - [preventing propagation] -> [toughness improvement].

[0030] Step S104: Input the target instruction and each first candidate evidence chain into the generation model to obtain the corresponding mechanism explanation text.

[0031] The target instruction is a natural language question or request posed by the user to the generative model, requiring explanation or response based on various first-candidate evidence chains. The generative model is a large language model, such as GPT-4, LLaMA, or ChatGLM. For example, based on all the first-candidate evidence chains in the input, it could explain how to improve the toughness of ultra-high performance concrete.

[0032] This disclosure employs a method for constructing a knowledge graph of ultra-high performance concrete (UHVPC) and using graph retrieval-enhanced generation for mechanism explanation. By acquiring multiple documents related to UHVPC and constructing a knowledge graph, it leverages the powerful semantic association and reasoning capabilities of knowledge graph technology to connect fragmented UHVPC knowledge. The UHVPC knowledge graph is used to perform adaptive length path retrieval on the target start and end nodes, obtaining at least one first candidate evidence chain to efficiently acquire the evidence chain corresponding to UHVPC knowledge. The target instruction and each first candidate evidence chain are then input into a generative model to obtain the mechanism explanation text output by the model. Multiple first candidate evidence chains provide the generative model with a precise and logically connected context, ensuring that the final generated mechanism explanation text has graph path support and high credibility. Thus, based on the constructed UHVPC knowledge graph, the evidence chain corresponding to UHVPC knowledge can be efficiently obtained, and then input into the generative model to obtain a highly credible mechanism explanation text for UHVPC knowledge, thereby achieving efficient acquisition of a highly credible mechanism explanation of UHVPC knowledge. This mechanism explanation text is traceable.

[0033] Preferably, multiple documents related to ultra-high performance concrete are obtained, including: based on a first strategy, selecting multiple initial journal articles from academic databases and multiple initial engineering reports from engineering report databases; wherein the first strategy includes the topic and / or abstract containing keywords corresponding to ultra-high performance concrete; based on a second strategy, selecting multiple target journal articles from all initial journal articles; the second strategy includes one or more of the following: publication time within a first preset time period, citation count greater than a preset citation value, and download count greater than a preset download count; based on a third strategy, selecting multiple target engineering reports from all initial engineering reports; wherein the third strategy includes publication time within a first preset time period and / or report content containing experimental information; the target journal articles and target engineering reports are used as documents related to ultra-high performance concrete to obtain multiple documents related to ultra-high performance concrete.

[0034] The first strategy broadens keyword coverage to encompass both cutting-edge journal articles and engineering reports reflecting practical engineering applications, thus balancing academic research and engineering practice and laying a multi-dimensional knowledge foundation. The second strategy applies objective indicators such as publication date, citation count, and download volume to a large pool of initial journal articles, precisely selecting core journal articles (i.e., target journal articles) with significant influence and timeliness in the field of ultra-high performance concrete. Considering the characteristics of engineering reports, a third strategy uses timeliness and / or "experimental information" as screening criteria, ensuring that selected engineering reports reflect the latest technological applications and verifiable quantitative data and specific methods, thereby enhancing their practical guidance value and avoiding mere empty project overviews. In this two-stage filtering mechanism, core literature with both academic depth and practical value can be extracted from a vast amount of journal articles and engineering reports, ensuring the quality of all visible ultra-high performance concrete knowledge graphs.

[0035] Preferably, the keywords corresponding to ultra-high performance concrete directly or indirectly describe the term "ultra-high performance concrete". For example, the Chinese and English names and abbreviations of ultra-high performance concrete (such as ultra-high performance concrete, UHPC, reactive powder concrete, Ultra-High Performance Concrete, etc.), and extended descriptions of ultra-high performance concrete (such as ultra-high performance fiber reinforced concrete, UHPFRC, environmentally friendly UHPC, etc.).

[0036] Preferably, a knowledge graph of ultra-high performance concrete is constructed based on multiple documents related to ultra-high performance concrete, including: extracting multiple basic sentence elements from various documents related to ultra-high performance concrete, and performing quality screening and terminology standardization on each basic sentence element to obtain multiple target sentence elements; using a target large language model to extract entities, attributes, and relations from each target sentence element, and generating multiple Cypher statements based on the entities, attributes, and relations corresponding to each target sentence element; and constructing a knowledge graph of ultra-high performance concrete based on each Cypher statement.

[0037] In this way, by extracting multiple basic sentence units from various literatures related to ultra-high performance concrete, and performing quality screening and terminology standardization, high-quality target sentence units are obtained. Then, the target large language model is called to extract entities, attributes, and relationships from the basic sentence units, and directly generate Cypher statements that can be executed in graph databases, so as to efficiently construct a knowledge graph of ultra-high performance concrete.

[0038] Preferably, multiple basic sentence elements are extracted from various documents related to ultra-high performance concrete, including: using a word segmenter to segment the continuous text stream in the documents into independent sentences to obtain multiple basic sentence elements. The basic sentence elements can be sentences or paragraphs.

[0039] Preferably, quality screening and terminology standardization are performed on each basic sentence element to obtain multiple target sentence elements, including: taking the set of all basic sentence elements as the basic sentence element set; performing a two-stage quality screening on each sentence element within the basic sentence element set, and using the screened basic sentence element set as the sentence element set to be standardized; performing terminology standardization on each sentence element within the sentence element set to be standardized to obtain the target sentence element set; and using each sentence element within the target sentence element set as the target sentence element to obtain multiple target sentence elements.

[0040] Preferably, a two-stage quality screening is performed on each sentence element in the basic sentence element set, including: traversing each sentence element in the basic sentence element set, identifying whether it triggers any rule in the pre-built rule set, and removing sentence elements that trigger any rule from the basic sentence element set, using the removed basic sentence element set as the initial screening sentence element set. Each sentence element in the initial screening sentence element set is input into a pre-trained quality classification model to obtain the quality level of each sentence element. Sentence elements with low quality levels are removed from the initial screening sentence element set, using the removed initial screening sentence element set as the sentence element set to be standardized.

[0041] Preferably, the rule set includes one or more of the following: sentence length rules (i.e., the sentence length of the sentence element is less than the first preset sentence length, or greater than the second preset sentence length), format error rules (i.e., the sentence element does not contain any verbs or subjects (based on simple grammar rules), or the sentence element does not contain key verbs), invalid character rules (i.e., the sentence element contains garbled characters or a large number of meaningless symbols), sensitive information rules (i.e., the sentence element contains phone numbers, email addresses, URLs, etc.), specific blacklisted word rules (i.e., the sentence element contains insulting or advertising words), citation rules (i.e., the sentence element contains citation identifiers, and the citation identifiers are in bracketed number format (e.g., [1], [3-5], etc.), bracketed author year format (e.g., (Smith et al., 2020), (Zhang and Lee, 2018) etc.), and citation identifiers (e.g., cited from...)).

[0042] Preferably, the quality classification model is obtained as follows: Multiple basic sentence elements are selected from the basic sentence element set as the original dataset. Each basic sentence element in the original dataset is labeled with a quality level: low, medium, or high, to obtain the target dataset (i.e., the sample features are the text content of a single basic sentence element, and the sample labels are the quality level corresponding to that basic sentence element (i.e., "high," "medium," and "low" labels)). ALBERT is fine-tuned using the target dataset, and the fine-tuned ALBERT is used as the quality classification model.

[0043] Preferably, terminology standardization is performed on each sentence element within the set of sentence elements to be standardized to obtain the target sentence element set. This includes: constructing a standardized dictionary; wherein the standardized dictionary includes standardized and non-standardized words. Chinese word segmentation tools (such as Jieba, HanLP, LTP, etc.) are used to segment and tag each sentence element within the set of sentence elements to be standardized; a NER model is used to identify entities in the segmented and tagged sentence elements to extract noun phrases as candidate terms. For each candidate term, a search is performed in the standardized dictionary. If a matching non-standardized word is found in the standardized dictionary, the candidate term is replaced with the corresponding standard term in the original sentence element (i.e., the sentence element within the set of sentence elements to be standardized) to obtain the target sentence element after terminology standardization. The set of all target sentence elements is taken as the target sentence element set.

[0044] Preferably, the entity, attribute, and relation in each target sentence element are extracted using the target large language model, and multiple Cypher statements are generated based on the entity, attribute, and relation corresponding to each target sentence element. This includes: extracting the entity, attribute, and relation in each target sentence element using the target large language model, and generating multiple candidate Cypher statements based on the entity, attribute, and relation corresponding to each target sentence element; forming a normalized mapping rule set based on all candidate Cypher statements; performing concept merging processing on the words in each candidate Cypher statement based on the normalized mapping rule set, and performing redundancy removal and updating on each candidate Cypher statement after processing to obtain multiple Cypher statements.

[0045] In this way, a standardized vocabulary is formed based on the words in the candidate Cypher statements, which can standardize the diverse terminology and merge and deduplicate words with the same semantics, thereby refining the Cypher statements and improving the quality of the ultra-high performance concrete knowledge graph.

[0046] Preferably, the entity, attribute and relation in each target sentence element are extracted using the target large language model, including: using a prompt word template of "instruction-input-output" to call the target large language model to extract the entity, attribute and relation in each target sentence element.

[0047] For example, the prompt template is: Please extract entities, attributes, and relationships from the following UHPC professional text and output them in JSON format: Text: {input_sentence} Output requirements: - Entity type: Material, composition, performance, process, testing standards - Relationship type: Includes performance, influencing factors, components, processes used, and standards followed. - Attributes include: numerical value, unit, condition, etc.

[0048] Preferably, the large language model is GPT-4, ChatGLM, etc. The target large language model is determined as follows: Using 50-100 target sentence units, the accuracy of each large language model in extracting entities, attributes, and relations is evaluated and used as the extraction accuracy. If at least one large language model has an extraction accuracy greater than or equal to a preset accuracy, the large language model with the highest extraction accuracy is selected as the target large language model. If the extraction accuracy of all large language models is less than the preset accuracy, each large language model is fine-tuned until at least one fine-tuned large language model has an extraction accuracy greater than or equal to the preset accuracy; the fine-tuned large language model with the highest extraction accuracy is selected as the target large language model.

[0049] Understandably, fine-tuning large language models is a relatively mature technique, so we will not go into details here.

[0050] Preferably, a normalized mapping rule set is formed based on all candidate Cypher statements, including: extracting term elements from all candidate Cypher statements, whereby term elements include at least one of node labels, relation types, attribute keys, and attribute values; performing synonym and near-synonym analysis on the extracted term elements to generate the normalized mapping rule set. The normalized mapping rule set stores the correspondence between normalized term elements and their corresponding normalized term elements.

[0051] Preferably, based on the normalized mapping rule set, the words in each candidate Cypher statement are subjected to concept merging processing, including: based on the normalized mapping rule set, each candidate Cypher statement is rewritten, and the non-normalized term elements in it are replaced with the corresponding normalized term elements to obtain each candidate Cypher statement after processing.

[0052] Preferably, the processed candidate Cypher statements are updated to remove redundancy, resulting in multiple Cypher statements. This includes: parsing the processed candidate Cypher statements to identify node operation instructions pointing to the same graph database node and relation operation instructions pointing to the same graph database relation; merging multiple node operation instructions pointing to the same graph database node and multiple relation operation instructions pointing to the same graph database relation; and generating and outputting the de-redundant Cypher statements based on the merging results to obtain multiple Cypher statements.

[0053] In this way, through concept merging and redundancy removal updates, the resulting Cypher statements are more concise in form, avoiding entity fragmentation and significantly improving query accuracy and interpretability. Furthermore, concept merging and normalization prevent entity and relationship fragmentation, improving graph quality and maintainability.

[0054] Preferably, a knowledge graph of ultra-high performance concrete is constructed based on each Cypher statement, including: importing each Cypher statement into a graph database (such as Neo4j) to form a queryable and visualized knowledge graph of ultra-high performance concrete.

[0055] Preferably, a method for constructing an ultra-high performance concrete knowledge graph and explaining the mechanism of graph retrieval-based enhanced generation further includes: having domain experts review the generated Cypher statements, removing redundancy, eliminating ambiguity, and correcting inconsistencies, thus implementing a human-machine system to ensure knowledge credibility. In this way, directly generating Cypher statements through LLM combined with human-machine collaboration significantly reduces graph construction costs and improves accuracy and coverage.

[0056] To facilitate understanding, this disclosure provides two examples to illustrate the mechanism of automated mapping and map-driven mapping. Among them, Example 1: Automated Knowledge Graph Construction: Collect several high-quality literature and engineering data related to ultra-high performance concrete, perform basic sentence elementization and two-stage quality screening on the text, and rewrite the retained basic sentence elements using terminology standardization; input the basic sentence elements into the target large language model to generate entity-relationship-attribute Cypher statements, then generate a unified vocabulary by merging synonyms / near-synonyms and regenerate a globally consistent Cypher; import the standardized statements into Neo4j to obtain an ultra-high performance concrete knowledge graph containing multiple types of entities (materials, processes, performance, mechanisms, etc.) and multiple types of relationships (influence, promotion, inhibition, participation in reactions, etc.); expert review is introduced throughout the process to eliminate ambiguity and inconsistency.

[0057] Example 2: Graph-driven mechanism explanation: For the problem of "doping with glass powder → compressive strength", start and end nodes are set, and variable-length path retrieval is used to obtain the evidence chain of all intermediate nodes and relationships, such as "volcanic ash reaction - filling effect - dilution effect", etc. The path and edge attributes (such as evidence source, confidence level, and scope of application) are structured and exported and input into the generative model to obtain traceable mechanism explanation text and applicable condition prompts (such as replacement rate range, maintenance system, etc.).

[0058] Preferably, a method for constructing a knowledge graph of ultra-high performance concrete and explaining the mechanism of graph retrieval-based enhanced generation further includes: calculating the path directness score, source authority score, scenario adaptation score, and solid waste reuse score for each first candidate evidence chain; wherein, the path directness score represents the directness of the first candidate evidence chain from the target start node to the target end node; the source authority score represents the academic credibility of the original knowledge source on which the first candidate evidence chain relies; the scenario adaptation score represents the matching length between the first candidate evidence chain and the user's demand scenario; the solid waste reuse score represents the value score of the first candidate evidence chain in terms of industrial solid waste resource reuse; and the path directness score is calculated according to the target end node. The system ranks each first candidate evidence chain based on its score, source authority score, scenario adaptability score, and solid waste reuse score, obtaining lists of path rankings, source authority rankings, scenario adaptability rankings, and solid waste reuse rankings. For each first candidate evidence chain, the system compares its ranking with all other first candidate evidence chains in each of these lists, determining the number of times the current first candidate evidence chain ranks higher. Based on the number of times each first candidate evidence chain ranks higher, the system determines the target evidence chain. Finally, the system determines the target design scheme based on the target evidence chain.

[0059] In this way, by comparing the ranking of each first candidate evidence chain with other first candidate evidence chains in the path ranking list, source authority ranking list, scenario adaptability ranking list, and solid waste reuse ranking list, the number of times each first candidate evidence chain ranks higher is obtained, and the target evidence chain is determined based on this. This ranking comparison method only cares about the relative order of each first candidate evidence chain compared with other first candidate evidence chains, which can avoid the differences caused by different dimensions. This ensures that the obtained target evidence chain is the first candidate evidence chain with better overall performance in a global comparison. It is the optimal solution that wins after sufficient comparison across multiple dimensions. Then, the target design method is determined based on the target evidence chain, making the target design method more credible, more logical, and more targeted.

[0060] In some embodiments, the path directness score characterizes the directness of the first candidate evidence chain from the target start node to the target end node. The path directness score can be determined by one or more of the following: path length, the relational strength (i.e., weight) of intermediate nodes, and path redundancy. For example, a shorter path length results in a higher path directness score. Alternatively, the path directness score can be obtained by weighted summing of the path length score, the relational strength (i.e., weight) score of intermediate nodes, and the path redundancy score. Here, a shorter path length results in a higher path length score; higher weights of intermediate nodes result in a higher relational strength score; and a greater number of irrelevant nodes in the path result in a lower path redundancy score.

[0061] In some embodiments, the source authority score characterizes the academic credibility of the original knowledge source upon which the first candidate chain of evidence relies. The source authority score can be determined by factors such as the source documents in the first candidate chain of evidence and the number of times those source documents have been cited in a recent period. The more authoritative the source document (authoritativeness can be determined through expert experience or publication level), the higher the source document score. If there are multiple source documents, the average score is calculated. The more times the source document has been cited in a recent period, the higher the citation score. The source authority score is obtained by weighted summing of the higher source document score and the citation score. The "recent period" can be such as the last 3 years, the last 1 year, etc.

[0062] In some embodiments, the scenario fit score represents the matching length between the first candidate evidence chain and the user's required scenario. If the user does not specify the required scenario, the scenario fit score can be 0. If the user specifies the required scenario, it is queried whether each first candidate evidence chain is applicable to the user's required scenario. If applicable, a first score is assigned to the applicable first candidate evidence chain; if not applicable, a second score is assigned to the inapplicable first candidate evidence chain; if it is unclear whether it is applicable, a third score is assigned to the first candidate evidence chain whose applicability is unclear. Wherein, the first score > the third score > the second score.

[0063] In some embodiments, the solid waste reuse score represents the value score of the first candidate evidence chain in terms of industrial solid waste resource reuse. The solid waste reuse score can be determined by one or more of the following: solid waste replacement rate, solid waste type fit, technological maturity, and environmental value. For example, a higher solid waste replacement rate results in a higher replacement score; a higher fit score is obtained if the solid waste type matches typical industrial by-products of the target area (e.g., fly ash promoted in power plant areas); a higher technological score is obtained if the solid waste utilization technology is in the laboratory stage or has mature engineering applications; and a higher value score is obtained if the potential estimated value of the carbon emission reduction, energy consumption reduction, and other benefits brought by the first candidate evidence chain is greater. The solid waste reuse score is then determined by weighted summation of the replacement score, fit score, technological score, and value score.

[0064] Understandably, each first candidate chain of evidence is ranked according to its direct path score, source authority score, scenario adaptability score, and solid waste reuse score: the higher the direct path score of a first candidate chain of evidence, the higher its ranking in the path ranking list; the higher the source authority score of a first candidate chain of evidence, the higher its ranking in the source authority ranking list; the higher the scenario adaptability score of a first candidate chain of evidence, the higher its ranking in the scenario adaptability ranking list; and the higher the solid waste reuse score of a first candidate chain of evidence, the higher its ranking in the solid waste reuse ranking list.

[0065] To facilitate understanding, and in conjunction with the ranking list, this disclosure provides an example of the number of times each first candidate chain of evidence is ranked higher: Ranking List Comparing the first candidate evidence chain 1 with the first candidate evidence chain 2, the first candidate evidence chain 1 ranks higher than the first candidate evidence chain 2 in the path ranking list, the source authority ranking list, and the solid waste reuse ranking list, and it ranks higher than the first candidate evidence chain 2 3 times. Comparing the first candidate evidence chain with the first candidate evidence chain 3, the first candidate evidence chain 1 ranks higher than the first candidate evidence chain 3 in the path ranking list and the source authority ranking list, and it ranks higher than the first candidate evidence chain 3 2 times. Therefore, the first candidate evidence chain 1 ranks higher 3 times + 2 times, for a total of 5 times.

[0066] Similarly, when comparing the first candidate evidence chain 2 with the first candidate evidence chain 1, the number of times the first candidate evidence chain 2 ranks higher than the first candidate evidence chain 1 is 1; when comparing the first candidate evidence chain 2 with the first candidate evidence chain 3, the number of times the first candidate evidence chain 2 ranks higher than the first candidate evidence chain 3 is 2; therefore, the number of times the first candidate evidence chain 2 ranks higher is 1 + 2, for a total of 3.

[0067] Similarly, when comparing the first candidate evidence chain 3 with the first candidate evidence chain 1, the number of times it ranks higher than the first candidate evidence chain 1 is 2; when comparing the first candidate evidence chain 3 with the first candidate evidence chain 2, the number of times it ranks higher than the first candidate evidence chain 2 is 2; therefore, the number of times the first candidate evidence chain 3 ranks higher is 2 + 2, for a total of 4.

[0068] Preferably, the target evidence chain is determined based on the number of times each first candidate evidence chain ranks higher, including: taking the number of times each first candidate evidence chain ranks higher as the candidate number and taking the largest candidate number as the target number; if the target number corresponds to only one first candidate evidence chain, then the first candidate evidence chain corresponding to the target number is taken as the target evidence chain; and if the target number corresponds to multiple first candidate evidence chains, then each first candidate evidence chain corresponding to the target number is taken as a second candidate evidence chain; the evaluation score of each second candidate evidence chain is determined based on the path direct score, source authority score, scenario adaptation score, and solid waste reuse score; and the second candidate evidence chain with the highest evaluation score is taken as the target evidence chain.

[0069] By comparing the highest-ranking occurrences and selecting the highest one as the target occurrence, we can quickly pinpoint the target occurrence. If the target occurrence corresponds to only one first-candidate evidence chain, it means that this first-candidate evidence chain is the optimal solution after sufficient comparison across multiple dimensions. Using this as the target evidence chain makes it the globally optimal evidence chain. If the target occurrence corresponds to multiple first-candidate evidence chains, it means that each of these first-candidate evidence chains (i.e., second-candidate evidence chains) is an optimal solution. In this case, by calculating the evaluation scores of each second-candidate evidence chain, we can select a more suitable target evidence chain from among the multiple optimal solutions. This allows us to quantify the selection of the target evidence chain and determine the optimal target design scheme.

[0070] Preferably, determining the target design scheme based on the target evidence chain includes: identifying at least one core node in the target evidence chain; wherein, the core node is a non-target start and end node; determining the first parameter range of each core node based on the target evidence chain, and converting the target evidence chain into a candidate design scheme based on each first parameter range; optimizing the candidate design scheme to determine the target design scheme.

[0071] In this way, by identifying the core nodes in the target evidence chain and determining the first parameter space of each core node, candidate design schemes are generated. The target design scheme is then determined by optimizing the candidate design schemes. This allows for improvements to the candidate design schemes, thereby increasing the efficiency of scheme design in the field of ultra-high performance concrete (UHPC).

[0072] Preferably, identifying at least one core node in the target evidence chain includes: identifying at least one core node in the target evidence chain based on directed relationships such as "enhancement / suppression" within the target evidence chain. This guides variable selection, mix design optimization, and other methods to determine the target design scheme.

[0073] Preferably, identifying at least one core node in the target evidence chain includes: identifying parameter tuning nodes among the intermediate nodes of the target evidence chain; matching each parameter tuning node with knowledge nodes in a preset knowledge base, designating successfully matched parameter tuning nodes as first nodes, and unmatched parameter tuning nodes as second nodes; calculating the importance of each second node, and filtering based on each importance, designating each filtered second node as a third node; and determining at least one core node in the target evidence chain based on each first node and each third node.

[0074] In this context, a parameter-tuning node represents a node whose parameters can be optimized and modified. That is, the edge attributes of this node are a parameter space, rather than fixed values.

[0075] Optimizing the parameters of all nodes in the target evidence chain would be too labor-intensive. Therefore, by matching each node with knowledge nodes in the knowledge base, a successful match indicates that the node is important and its parameter tuning value is high, allowing for the rapid identification of the first node. If a match fails, the unmatched node is designated as the second node. The importance of each second node is then determined and filtered to identify those that meet the importance criteria as third nodes, further determining the third nodes with the highest parameter tuning value. This approach, by identifying the first and third nodes, balances filtering efficiency and quality, resulting in at least one core node in the target evidence chain.

[0076] Preferably, the importance of each second node is calculated, and a selection is made based on the importance, with the selected second nodes being designated as third nodes. This includes: calculating the frequency of each second node's occurrence in all first candidate evidence chains, and using the frequency of occurrence as the importance of each second node. Second nodes with an importance greater than a preset importance threshold are designated as nodes.

[0077] Preferably, determining at least one core node in the target evidence chain based on each first node and each third node includes: treating all first nodes and all third nodes as core nodes to obtain at least one core node in the target evidence chain. Alternatively, selecting at least one node from all first nodes and all third nodes as a core node to obtain at least one core node in the target evidence chain. No limitation is imposed here.

[0078] Preferably, determining the first parameter range of each core node based on the target evidence chain includes: determining the edge attributes of each core node based on the target evidence chain; and determining the first parameter space of the corresponding core node based on each edge attribute.

[0079] Preferably, the candidate design scheme includes the first parameter space of each core node. Optimizing the candidate design scheme to determine the target design scheme includes: collecting multiple experimental sample data corresponding to the target evidence chain for each first parameter interval within the candidate design scheme, and optimizing the first parameter interval of each core node based on the multiple experimental sample data to obtain the second parameter interval of each core node; and converting the target evidence chain into a design scheme based on the second parameter space of each core node to obtain the target design scheme.

[0080] Since the first parameter space determined based on the target evidence chain is the preferred parameter space given by the knowledge graph, the feasibility and effectiveness of the candidate design scheme remain to be discussed when it is converted into a candidate design scheme. Therefore, by collecting relevant experimental sample data and using it to optimize the candidate design scheme, improvements can be made to the candidate design scheme, thereby improving the scheme design efficiency in the UHPC field.

[0081] Understandably, the candidate design scheme is actually a design scheme that converts the target evidence chain and the first parameter space of its core nodes into natural language or structured data.

[0082] To facilitate understanding, this disclosure provides an example illustrating the process of determining the target design scheme: In a glass powder-containing system, based on spectral retrieval, the branches and synergistic mechanisms that enhance strength are identified, and the ternary system path of "glass powder + slag" is found as the target evidence chain. Spectral evidence indicates that high CaO content (i.e., the core node) promotes secondary hydration reaction and densification process. Based on this, a list of candidate schemes is generated, and the parameter range of CaO content is constrained and sorted in combination with existing experimental data, outputting an executable design suggestion, which is the target design scheme.

[0083] In some embodiments, the priority of each core node can be calculated first, and when the first parameter range of each core node is tuned, the needs of the high-priority core nodes are satisfied first.

[0084] In other embodiments, when optimizing each first parameter interval, the first parameter interval of each core node can be optimized based on a genetic algorithm and multiple experimental sample data.

[0085] Preferably, combined with Figure 2 As shown, Figure 2A flowchart illustrating a method for updating a knowledge base is provided. A mechanism explanation method for constructing a knowledge graph of ultra-high performance concrete and enhancing its generation based on graph retrieval also includes updating the knowledge base through the following methods: Step S201: Calculate the matching activity of knowledge nodes in the current knowledge base, and downgrade knowledge nodes with matching activity less than the first threshold to the cache.

[0086] Step S202: Match each third node with the cache nodes in the cache library. If the match fails, add the third node that failed to match to the cache library.

[0087] Step S203: After a first preset time interval, update the cache activity of each cache node in the cache library and update the total time for each cache node to enter the cache library.

[0088] Step S204: Upgrade cache nodes with a total duration ≥ the second preset duration and a cache activity ≥ the second threshold to the knowledge base to obtain the updated knowledge base.

[0089] Among these conditions, the initial cache activity level is less than the second threshold, and the first preset duration is less than the second preset duration.

[0090] In this way, by downgrading, knowledge nodes with low matching activity can be promptly deleted from the knowledge base, updating the knowledge base accordingly. Then, by matching each third-party node with cache nodes in the cache repository, and updating the cache repository with third-party nodes that fail to match, the cache nodes in the cache repository are updated promptly. By updating the cache activity and total duration of each cache node, cache nodes that meet the upgrade conditions are upgraded to the knowledge base, achieving incremental updates to the knowledge base. This creates a dynamic cycle of "knowledge base downgrading → cache repository temporary storage → third-party node replenishment → cache repository upgrade → knowledge base backflow," ensuring the "core" nature of the knowledge base (i.e., retaining only knowledge nodes with high matching activity) while preventing potentially valuable nodes from being directly discarded through the cache repository.

[0091] Understandably, the knowledge nodes in the initial knowledge base are typically determined based on expert experience. An initial matching activity level is assigned to cache nodes newly added to the knowledge base. This initial matching activity level is greater than a first threshold. The total time a cache node spends in the cache is the time interval between the initial entry time of that cache node and the current time.

[0092] In some embodiments, calculating the matching activity of knowledge nodes in the current knowledge base includes: for each knowledge node in the current knowledge base, its matching activity = matching activity at the end of the previous period - basic decay + matching gain. If it is a knowledge node newly added to the knowledge base (e.g., a node added from a cache or in response to expert experience), its matching activity at the end of the previous period is a preset initial matching activity. The basic decay is a preset value.

[0093] In some embodiments, the matching gain is determined by the number of successful matches between the parameter-tuning node and the knowledge node within the current period, and the more successful matches within the current period, the greater the matching gain. For example, matching gain = base score for a single successful match × total number of successful matches in the current period. Or, , For candidate gain, This is the preset upper limit of gain. Candidate gain values ​​are calculated using the following method: Where N represents the number of successful matches in this period. , and The base score for a single successful match.

[0094] Preferably, updating the cache activity of each cache node in the cache repository includes: using the current cache activity of each cache node in the cache repository as the cache activity before the update, and obtaining the time interval between each cache node and the last successful match; determining the update degree corresponding to each time interval. Based on the cache activity of each cache node before the update and the corresponding update degree, the cache activity of each cache node after the update is determined, so as to update the cache activity of each cache node in the cache repository.

[0095] In this way, by determining the time interval between the last successful match on each cache node, the update degree of each cache node is determined, and this update degree is directly related to the time interval. Then, the current cache degree of the cache node is updated based on the update degree and the cache activity before the update, thereby realizing dynamic updating of cache activity.

[0096] It's understandable that the time interval between a cache node's last successful match and its current matching time refers to the time interval between the cache node's last successful match with a third node. If a cache node is a knowledge node that has just been downgraded to the cache repository, or a third node that was just added to the cache repository after failing to match with a cache node already in the cache repository, then its current cache activity is the initial cache activity. This initial cache activity is greater than the second threshold.

[0097] Understandably, the longer the time interval, the lower the corresponding update frequency. This scientifically reflects the characteristic that the timeliness of content on cached nodes gradually decreases over time.

[0098] For example, the update rate can be calculated using the following formula: The parameters of this formula can be determined based on expert experience. for The corresponding update rate, This is the time interval, which is the number of days between intervals.

[0099] Preferably, the cache activity of each cache node after the update is determined based on the cache activity and update rate of each cache node before the update, including: for each cache node, the sum of the cache activity and update rate before the update is used as the cache activity after the update, so as to determine the cache activity of each cache node after the update.

[0100] Preferably, a method for constructing an ultra-high performance concrete knowledge graph and explaining the mechanism of graph retrieval-based enhanced generation further includes: designating cache nodes with a total duration ≥ a second preset duration and a third threshold ≤ cache activity < the second threshold as pending nodes, and generating a list of pending nodes for review; sending the list of pending nodes to multiple reviewers, and determining the review status of each pending node in response to the review results provided by each reviewer; wherein the review status includes the number of times each review status has been applied, and the review status includes upgrade or deletion; upgrading each pending node to the knowledge base or deleting each pending node from the cache database based on the review status of each pending node; and deleting cache nodes with a total duration ≥ a second preset duration and a cache activity < the third threshold from the cache database.

[0101] If the total duration is greater than or equal to the second preset duration, and the third threshold is less than or equal to the cache activity level but less than the second threshold, it indicates that the cache node is relatively active. In this case, a list of nodes awaiting review is generated and sent to multiple reviewers. Each reviewer reviews and summarizes the results for each node on the list to obtain the review status. Based on the reviewers' votes on each node, it can be upgraded to the knowledge base or removed from the cache. This avoids the accidental deletion of important cache nodes even if their activity level does not directly meet the upgrade requirements, achieving human-machine collaboration and balancing the efficiency and quality of knowledge base updates. Conversely, cache nodes with a total duration greater than or equal to the second preset duration and a cache activity level less than the third threshold have low activity levels and are removed from the cache to promptly eliminate inactive cache nodes.

[0102] For example, if there are three pending nodes in the audit list: Pending Node A, Pending Node B, and Pending Node C, and this list is sent to three auditors: Auditor A, Auditor B, and Auditor C, the audit results from Auditor A are: Pending Node A: Upgrade; Pending Node B: Delete; Pending Node C: Upgrade. Auditor B's results are: Pending Node A: Upgrade; Pending Node B: Upgrade; Pending Node C: Delete. Auditor C's results are: Pending Node A: Delete; Pending Node B: Delete; Pending Node C: Upgrade. Therefore, the summary audit status for each pending node is: Pending Node A was upgraded 2 times (Auditor A and Auditor B). The number of deletions for pending node B is 1 (Auditor C). The number of upgrades for pending node B is 1 (Auditor B), and the number of deletions is 2 (Auditor A and Auditor C). The number of upgrades for pending node C is 2 (Auditor A and Auditor C), and the number of deletions is 1 (Auditor B).

[0103] Based on the review status of each pending node, each pending node can be upgraded to the knowledge base or deleted from the cache. Specifically, if the number of upgrades for each pending node is greater than 60% of the total number of reviewers, the pending node will be upgraded to the knowledge base; otherwise, it will be deleted from the cache.

[0104] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for constructing a knowledge graph of ultra-high performance concrete and explaining the mechanism of graph retrieval-based enhanced generation, characterized in that, include: Obtain multiple documents related to ultra-high performance concrete; Based on multiple literatures related to ultra-high performance concrete, a knowledge graph of ultra-high performance concrete is constructed. Based on the knowledge graph of ultra-high performance concrete, an adaptive length path retrieval is performed on the target origin and destination nodes to obtain at least one first candidate evidence chain corresponding to the target origin and destination nodes; The target instruction and each first candidate evidence chain are input into the generative model to obtain the corresponding mechanism explanation text; Also includes: Calculate the path directness score, source authority score, scenario adaptation score, and solid waste reuse score for each first candidate evidence chain. The path directness score represents the directness of the first candidate evidence chain from the target start node to the target end node; the source authority score represents the academic credibility of the original knowledge source on which the first candidate evidence chain relies; the scenario adaptation score represents the matching length between the first candidate evidence chain and the user's demand scenario; and the solid waste reuse score represents the value score of the first candidate evidence chain in terms of industrial solid waste resource reuse. The first candidate evidence chains are ranked according to the direct path score, the authoritative source score, the scenario adaptation score, and the solid waste reuse score, respectively, to obtain the path ranking list, the authoritative source ranking list, the scenario adaptation ranking list, and the solid waste reuse ranking list. For each first candidate evidence chain, compare the current first candidate evidence chain with the other first candidate evidence chains in the path ranking list, source authority ranking list, scenario adaptation ranking list and solid waste reuse ranking list, and determine the number of times the current first candidate evidence chain ranks higher. The target evidence chain is determined based on the number of times each first candidate evidence chain ranks higher. Determine the target design scheme based on the target evidence chain; The applicable score for a scenario is determined as follows: If the user does not specify the required scenario, the applicable score for that scenario is 0; if the user specifies the required scenario, it is checked whether each first candidate evidence chain is applicable to the user's required scenario. If applicable, the applicable first candidate evidence chain is assigned a first score value for the scenario applicable score; if not applicable, the inapplicable first candidate evidence chain is assigned a second score value for the scenario applicable score; if it is unclear whether it is applicable, the inapplicable first candidate evidence chain is assigned a third score value for the scenario applicable score, where the first score value > the third score value > the second score value.

2. The method according to claim 1, characterized in that, The acquisition of multiple documents related to ultra-high performance concrete includes: Based on the first strategy, multiple initial journal articles were selected from academic databases and multiple initial engineering reports were selected from engineering report databases; the first strategy included topics and / or abstracts containing keywords corresponding to ultra-high performance concrete. Based on the second strategy, multiple target journal articles are selected from all initial journal articles; the second strategy includes one or more of the following: publication time within a first preset time period, number of citations greater than a preset citation value, and number of downloads greater than a preset number of downloads. Based on the third strategy, multiple target engineering reports are selected from all initial engineering reports; the third strategy includes the publication time within a first preset time and / or the report content containing experimental information; By using target journal articles and target engineering reports as relevant literature for ultra-high performance concrete, multiple relevant documents for ultra-high performance concrete can be obtained.

3. The method according to claim 1, characterized in that, The aforementioned knowledge graph of ultra-high performance concrete is constructed based on multiple documents related to ultra-high performance concrete, including: Multiple basic sentence elements were extracted from various literatures related to ultra-high performance concrete, and quality screening and terminology standardization were performed on each basic sentence element to obtain multiple target sentence elements. The entity, attribute, and relation of each target sentence element are extracted using the target large language model, and multiple Cypher statements are generated based on the entity, attribute, and relation of each target sentence element. A knowledge graph of ultra-high performance concrete is constructed based on each Cypher statement.

4. The method according to claim 3, characterized in that, The process involves extracting entities, attributes, and relations from each target sentence element using a target large language model, and generating multiple Cypher statements based on the entities, attributes, and relations corresponding to each target sentence element, including: The entity, attribute, and relation of each target sentence element are extracted using the target large language model, and multiple candidate Cypher sentences are generated based on the entity, attribute, and relation of each target sentence element. Based on all candidate Cypher statements, a normalized mapping rule set is formed; Based on the normalized mapping rule set, the words in each candidate Cypher statement are merged into concepts, and the processed candidate Cypher statements are updated to remove redundancy, resulting in multiple Cypher statements.

5. The method according to claim 1, characterized in that, The determination of the target evidence chain based on the number of times each first candidate evidence chain ranks higher includes: The number of times each first candidate evidence chain ranks higher is taken as the candidate count, and the maximum candidate count is taken as the target count. If the target number corresponds to only one first candidate evidence chain, then the first candidate evidence chain corresponding to the target number is taken as the target evidence chain; if the target number corresponds to multiple first candidate evidence chains, then each first candidate evidence chain corresponding to the target number is taken as a second candidate evidence chain. The evaluation score for each second candidate evidence chain is determined based on the path direct score, source authority score, scenario adaptation score, and solid waste reuse score. The second candidate evidence chain with the highest evaluation score is selected as the target evidence chain.

6. The method according to claim 1, characterized in that, The determination of the target design scheme based on the target evidence chain includes: Identify at least one core node in the target evidence chain; wherein, the core node is not the target start or end node; The first parameter range of each core node is determined based on the target evidence chain, and the target evidence chain is converted into a candidate design scheme based on each first parameter range. The candidate design schemes are optimized to determine the target design scheme.

7. The method according to claim 6, characterized in that, At least one core node in the chain of evidence for identifying the target includes: Identify the parameter-tuning nodes in the middle nodes of the target evidence chain; Each parameter tuning node is matched with a knowledge node in a preset knowledge base. The parameter tuning node that matches successfully is designated as the first node, and the parameter tuning node that does not match successfully is designated as the second node. Calculate the importance of each second node, and filter them based on their importance, then select the selected second nodes as the third nodes. At least one core node in the target evidence chain is determined based on each first node and each third node.

8. The method according to claim 7, characterized in that, Also includes: Update the knowledge base in the following ways: Calculate the matching activity of knowledge nodes in the current knowledge base, and downgrade knowledge nodes with matching activity less than the first threshold to the cache library; Each third node is matched with the cache nodes in the cache library. If a match fails, the third node that fails to match is added to the cache library. At a first preset interval, update the cache activity of each cache node in the cache library and update the total time for each cache node to enter the cache library; Upgrade cache nodes with a total duration greater than or equal to the second preset duration and a cache activity level greater than or equal to the second threshold to the knowledge base to obtain the updated knowledge base.

9. The method according to claim 8, characterized in that, Also includes: Cache nodes with a total duration greater than or equal to the second preset duration and a third threshold less than or equal to the cache activity level less than the second threshold are designated as pending nodes, and a list of pending nodes is generated based on each pending node. The list of items pending review is sent to multiple reviewers, and the review status of each pending node is determined in response to the review results provided by each reviewer. The review status includes the number of times each review status has been applied, and the review status includes upgrade or deletion. Based on the review status of each pending node, upgrade each pending node to the knowledge base, or delete each pending node from the cache. Cache nodes with a total duration greater than or equal to the second preset duration and a cache activity level less than the third threshold will be removed from the cache library.

Images