A composite knowledge chain construction method for scientific research logic representation

Abstract

The application provides a composite knowledge chain construction method for scientific research logic characterization, and relates to the technical field of knowledge extraction. The method comprises the following steps: constructing a composite knowledge chain structure model by defining core knowledge components; constructing an instruction set of components and sub-elements, and fine-tuning a pre-trained large language model; processing scientific literature through the fine-tuned large language model to generate a knowledge extraction result; executing analysis of internal structural relationships of components and cross-component logical relationships; structurally organizing components, internal element relationships of components and cross-component logical relationships to generate a composite knowledge chain. The application can solve the technical problem of low accuracy of scientific research knowledge extraction in the prior art due to the difficulty of knowledge organization methods in deeply characterizing complex logical relationships and dynamic dependencies in scientific research activities. By deeply characterizing scientific research logic and automatically constructing a composite knowledge chain, the accuracy of knowledge extraction is improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge extraction technology, specifically to a method for constructing a composite knowledge chain oriented towards scientific research logical representation. Background Technology

[0002] Scientific literature, as the primary carrier of research results, records the complete process from problem formulation to solution seeking and result verification. However, existing technologies have several shortcomings in representing the inherent logic of scientific research activities, especially in the organization, extraction, and deep understanding of scientific knowledge. Scientific research activities involve multiple elements with complex logical relationships and interdependencies. Existing methods can usually only capture macro-level topic classifications and basic semantic relationships, making it difficult to deeply reveal the dynamic interactions, driving factors, and complex logical dependencies within scientific research activities. In recent years, knowledge representation technologies, represented by knowledge graphs, have been widely applied. Knowledge graphs store factual knowledge in the form of (head entity, relation, tail entity) triples, clearly revealing explicit relationships between entities. However, standard knowledge graphs have shortcomings in expressing complex, multi-faceted relationships. Hypergraphs, as a more flexible technology, connect multiple nodes through hyperedges, enabling the representation of multi-faceted relationships and providing richer knowledge representations. However, existing research based on hypergraphs focuses more on static multi-faceted associations, lacking in-depth representation of dynamic semantic roles (such as problems, methods, environments, and results) and their directed, dynamic logical flows within scientific research activities.

[0003] In summary, existing technologies suffer from the problem that the knowledge organization methods are unable to deeply represent the complex logical relationships and dynamic dependencies in scientific research activities, resulting in low accuracy in extracting scientific knowledge. Summary of the Invention

[0004] This application provides a method for constructing a composite knowledge chain oriented towards the logical representation of scientific research, which addresses the technical problem in existing technologies where the accuracy of scientific knowledge extraction is low because knowledge organization methods are unable to deeply represent the complex logical relationships and dynamic dependencies in scientific research activities.

[0005] In view of the above problems, this application provides a method for constructing a composite knowledge chain oriented towards scientific research logical representation. The method includes: constructing a composite knowledge chain structure model by defining core knowledge components; based on the composite knowledge chain structure model, constructing instruction sets for components and sub-elements, and fine-tuning a pre-trained large language model; processing scientific and technological documents using the fine-tuned large language model to generate knowledge extraction results; based on the knowledge extraction results, parsing the internal structural relationships and cross-component logical relationships of components to generate internal element relationships and cross-component logical relationships; and structurally organizing the components, the internal element relationships, and the cross-component logical relationships to generate a composite knowledge chain.

[0006] Optionally, the core knowledge components include at least a problem component, a method component, an environment component, and a result component.

[0007] Optionally, the sub-elements of the problem component include background information, research motivation, research gap, and research objectives; the sub-elements of the method component include theoretical basis, key technologies, operational steps, and tools used; the sub-elements of the environment component include experimental conditions, underlying assumptions, disciplinary background, and constraints; and the sub-elements of the results component include findings, experimental data, conclusions, evaluation, and discussion.

[0008] Optionally, the core knowledge components are hierarchically and sequentially associated with their internal elements to construct a structural relationship diagram; a set of logical relationship types for connecting the core knowledge components is preset, including driving relationships, constraint relationships, influence relationships, resolution relationships, derivation relationships, and support or challenge relationships; the structural relationship diagram and the set of logical relationship types are combined to perform a formal representation of the composite knowledge chain, generating the composite knowledge chain structure model.

[0009] Optionally, a hierarchical system of relationship types is introduced, including basic relationships, general research behaviors in a field, and relationships of specific research paradigms; by analyzing research literature samples from multiple fields through the hierarchical system, an interdisciplinary relationship sub-library is constructed to supplement the set of logical relationship types.

[0010] Optionally, the formal representation of a composite knowledge chain includes a structured object containing core knowledge components, their internal elements, and the relationships between components.

[0011] Optionally, the hierarchy, temporal sequence, and dependency relationships between elements within the same component in the knowledge extraction results are parsed to generate the relationships between elements within the component; a pre-trained classifier is invoked using preset prompts or instructions to determine whether there are predefined logical relationships between two or more extracted component elements, thereby generating the cross-component logical relationships.

[0012] Optionally, logical rules are defined to perform consistency verification and conflict resolution on the components, the relationships between elements within the components, and the logical relationships across components, generating a verification fusion result; based on the verification fusion result, the identified component elements are used as nodes and the relationships are used as edges to construct a component relationship graph; each element and relationship in the component relationship graph is assigned a confidence score to generate the composite knowledge chain.

[0013] Optionally, a multi-factor confidence assessment model is constructed; wherein the multi-factor confidence assessment model is constructed based on the output probability of a large language, the text support span, the matching correspondence with existing knowledge graphs, and the frequency of occurrence in multiple relevant documents; and a confidence score is assigned to each element and relation in the component relationship graph using the multi-factor confidence assessment model.

[0014] Optionally, the composite knowledge chain is stored in the form of a graph database, supporting the execution of scientific research question analysis, result tracking, relationship chain mining, and knowledge recommendation through a structured query language; wherein, the structured query language includes structural representations based on structural perspectives, temporal perspectives, or thematic perspectives.

[0015] The technical solution provided in this application has at least the following beneficial effects:

[0016] A composite knowledge chain structure model is constructed by defining core knowledge components. Based on this model, an instruction set for the components and sub-elements is built, and a pre-trained large language model is fine-tuned. The fine-tuned model then processes scientific and technological documents to generate knowledge extraction results. Based on these results, the internal structural relationships and cross-component logical relationships of the components are analyzed to generate internal element relationships and cross-component logical relationships. Finally, the components, their internal element relationships, and cross-component logical relationships are structured to generate a composite knowledge chain. In other words, by defining core knowledge components to construct a composite knowledge chain structure model and fine-tuning a pre-trained large language model to process scientific and technological documents, analyze the relationships between internal elements and cross-component logical relationships, and automatically construct a composite knowledge chain, the accuracy of knowledge extraction is improved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a composite knowledge chain construction method for scientific research logical representation proposed in this application.

[0019] Figure 2 This is a schematic diagram of the process for generating a composite knowledge chain in the composite knowledge chain construction method for scientific research logical representation proposed in this application. Detailed Implementation

[0020] This application provides a method for constructing composite knowledge chains oriented towards the logical representation of scientific research. This addresses the technical problem in existing technologies where the difficulty in deeply representing the complex logical relationships and dynamic dependencies in scientific research activities leads to low accuracy in knowledge extraction. By defining core knowledge components to construct a composite knowledge chain structure model and fine-tuning a pre-trained large language model, the method processes scientific literature, analyzes the relationships between elements within components and the logical relationships across components, and automatically constructs the composite knowledge chain, thereby improving the accuracy of knowledge extraction.

[0021] Examples, such as Figure 1 As shown, this application provides a method for constructing a composite knowledge chain oriented towards scientific research logical representation. Specifically, this method includes the following steps:

[0022] A composite knowledge chain structure model is constructed by defining core knowledge components.

[0023] Furthermore, this application also includes the following steps: the core knowledge component includes at least a problem component, a method component, an environment component, and a result component.

[0024] Furthermore, this application also includes the following steps: the sub-elements of the problem component include background information, research motivation, research gap, and research objectives; the sub-elements of the method component include theoretical basis, key technologies, operational steps, and tools used; the sub-elements of the environment component include experimental conditions, underlying assumptions, disciplinary background, and constraints; and the sub-elements of the results component include findings, experimental data, conclusions, evaluation, and discussion.

[0025] Furthermore, this application also includes the following steps: hierarchically and sequentially associating the internal elements of the core knowledge components to construct a structural relationship diagram; pre-setting a set of logical relationship types for connecting the core knowledge components, including driving relationships, constraint relationships, influence relationships, resolution relationships, derivation relationships, and support or challenge relationships; combining the structural relationship diagram and the set of logical relationship types to perform a formal representation of the composite knowledge chain, and generating the composite knowledge chain structure model.

[0026] Furthermore, this application also includes the following steps: introducing a hierarchical system of relation types, including basic relations, general research behaviors in a field, and specific research paradigm relations; and constructing an interdisciplinary relation sub-library by analyzing research literature samples from multiple fields through the hierarchical system to supplement the set of logical relation types.

[0027] Furthermore, this application also includes the following steps: the formal representation of the composite knowledge chain includes a structured object containing core knowledge components and their internal elements and relationships between components.

[0028] Specifically, core knowledge components are defined as the basic units that constitute a composite knowledge chain, including problem components, methodological components, environmental components, and outcome components. The problem component is the core issue to be explored or solved in the research activity, including sub-elements such as background information, research motivation, research gaps, specific research objectives, or hypotheses to be verified. The methodological component comprises the theories, technologies, processes, algorithms, and experimental designs used to solve a specific problem, including sub-elements such as theoretical foundations, key technologies or algorithms, specific operational steps or processes, and tools / software / datasets / experimental materials used. The environmental component refers to the specific conditions, constraints, or context in which the research activity takes place, including sub-elements such as experimental conditions / parameter settings, research assumptions, specific application fields or disciplinary backgrounds, and limitations or constraints faced by the research. The outcome component comprises the findings, conclusions, data, and phenomena obtained by applying specific methods in a specific environment for a specific problem, including sub-elements such as main findings or observed phenomena, specific data or indicators obtained from experiments, conclusions or inferences drawn, evaluation or verification of results, and discussion and interpretation of results. Sub-elements refer to the specific constituent elements of each component, used to refine the specific content of that component.

[0029] When constructing a composite knowledge chain, each core knowledge component (such as a problem component, method component, environment component, and result component) contains multiple sub-elements, and these sub-elements have hierarchical and sequential relationships. Hierarchical structure refers to the structural relationships between sub-elements within a core knowledge component, such as the sequential relationship of method steps. Sequential association refers to the arrangement of sub-elements within a component according to a certain rule or order, establishing sequential relationships between them. For example, the sub-elements of the problem component (such as background information, research motivation, and research gaps) logically progress layer by layer. Background information provides background support for research motivation, and research gaps point out deficiencies in the background, thus deriving the research objective. The sub-elements of the method component (such as operational steps) need to be executed sequentially; therefore, their sequential association in the method component manifests as the sequential relationship of steps. By performing hierarchical and sequential association on the internal elements of the core knowledge component, a structural relationship diagram is constructed and represented by a directed acyclic graph (DAG). For example, suppose that in a chemical experiment, the steps of the method components include: preparing reagents (step 1), mixing reagents (step 2), heating the reaction (step 3), cooling (step 4), and measuring the results (step 5). The sequential relationship between these steps can be represented by a DAG, where step 1 points to step 2, step 2 points to step 3, and so on.

[0030] A pre-defined set of logical relationship types for connecting core knowledge components defines various logical relationships between components, such as driving relationships, constraint relationships, influence relationships, resolution relationships, derivation relationships, and support or challenge relationships. Driving relationships indicate how one component (e.g., a problem component) drives the selection and design of another component (e.g., a method component); for example, a research problem might drive a specific experimental method. Constraint relationships indicate the constraint effect of one component (e.g., an environmental component) on other components (e.g., method or result components); for example, experimental conditions might constrain the selection of experimental methods or experimental results. Influence relationships indicate the impact of one component (e.g., a method or environmental component) on the result component; for example, a certain experimental method or specific experimental conditions have varying degrees of influence on the experimental results.

[0031] A solution relationship refers to how a method component addresses the problem described by the problem component; it typically describes the degree to which the method component solves the problem component. A derivation relationship indicates that the outcome component of a study leads to a new problem component. A support or challenge relationship indicates that the outcome component of one study supports / challenges the hypothesis of another problem component.

[0032] By using a pre-defined set of logical relationship types, the interactions and dependencies between components in scientific research activities are revealed. Multiple relationships in the set can be represented by labeled directed edges. A hierarchical relationship type system is introduced, expanding the original relationship types to three levels: basic relationships, domain-wide general research behaviors, and specific research paradigm relationships. An interdisciplinary relationship sub-base is constructed by analyzing scientific literature from multiple fields (such as biology, physics, and computer science). The expressive power of composite knowledge chain structure models, such as control group-processing group, negation verification, and inference chain support, is enhanced.

[0033] The hierarchical system of relation types is a structure for classifying logical relationships, divided into multiple levels, from the most basic general relations to domain-specific relations, including basic relations, domain-general research behavior relations, and research paradigm-specific relations. Basic relations are the most fundamental and common types, such as driving relations and constraint relations, applicable to any field. Domain-general research behaviors are research behaviors and patterns specific to a particular discipline, common in research across different fields but not limited to a specific paradigm, such as experimental verification relations and theoretical derivation relations. Research paradigm-specific relations are relations unique to a specific research method or pattern (such as controlled experiments or negative verification), such as control group-treatment group relations and negative verification relations.

[0034] By analyzing research literature samples from multiple fields, such as biology, physics, and computer science, core logical relationships within these fields are extracted and organized to establish an interdisciplinary relationship sub-library containing relationship samples from research literature across different disciplines. Combining the hierarchical system with the interdisciplinary relationship sub-library further enhances the expressive power of the composite knowledge chain model. Particularly when expressing complex research patterns such as control-treatment groups, negative verification, and chain-based inference support, the composite knowledge chain structure model can more accurately and flexibly represent the logical relationships of these patterns. In other words, by supplementing the existing set of logical relationship types with new relationship types from the interdisciplinary relationship sub-library, the entire set becomes more comprehensive and detailed, covering a wider range of research logic expression needs. This ensures that the composite knowledge chain structure model can not only express general research logic but also accurately capture specific logical patterns under different disciplines and research methods.

[0035] Composite Knowledge Chain CKC i It can be formally represented as a structured object containing core components, their internal elements, and the relationships between components. For example, it can be represented as an extended graph structure G = (V, L). V ,A,L A ), where V is the set of knowledge element nodes (from the problem component, method component, environment component, and result component), L V A is the type label of the node; A is the set of relation edges between features, and L is the type label of the node. A It is the type label for the edge.

[0036] More specifically, a composite knowledge chain can be represented as an ordered tuple: CKC i =(P i M i E i ,R i ,R intra ,R inter Metadata), where P i M i E i ,R i These represent instances of the problem component, method component, environment component, and result component in the knowledge chain, respectively; R intra A structural relationship diagram representing the set of relationships between elements within each component; R inter It represents a collection of logical relationship types between components; Metadata contains metadata such as literature sources and time.

[0037] A formal representation of a composite knowledge chain is achieved by combining a structural relationship diagram and a set of logical relationship types. This involves formally representing each component, its internal element relationships, and the logical relationships between components to obtain a composite knowledge chain structural model. The goal of this formal representation is to present all core components of research activities, their internal element relationships, and the logical relationships between components in the form of structured objects, using ordered tuples, graph models, or other formal expression methods to represent the composite knowledge chain.

[0038] The formal representation of a composite knowledge chain includes a structured object containing core knowledge components, their internal elements, and the relationships between components. In other words, a composite knowledge chain will contain the hierarchical relationships of all core knowledge components and their internal elements, and connect different components through logical relationships (such as driving relationships, influence relationships, etc.).

[0039] The composite knowledge chain is a structured knowledge representation model that uses Problem-Method-Environment-Result (PMER) as its core component and explicitly represents the complex logical relationships such as driving forces, constraints, and influences between components. By constructing a structural relationship diagram and a pre-defined set of logical relationship types, it deeply represents the complex logical relationships and dynamic dependencies in scientific research activities. The structural relationship diagram clarifies the hierarchical and sequential relationships between elements within a component, while the set of logical relationship types clarifies the core logical relationships between components. The composite knowledge chain can accurately demonstrate the relationships between various stages and elements in scientific research activities, and how they affect the final result, providing clear structured knowledge.

[0040] Based on the aforementioned composite knowledge chain structure model, an instruction set for components and sub-elements is constructed, and the pre-trained large language model is fine-tuned. The fine-tuned large language model is then used to process scientific and technological documents to generate knowledge extraction results.

[0041] Specifically, based on a composite knowledge chain structure model, a detailed and structured instruction set is designed for the four core components—problem component, method component, environment component, and result component—and their sub-elements. This instruction set is then fine-tuned on the pre-trained large language model to enable it to accurately identify and extract text fragments or key information points related to each component. When generating the extraction results, the large language model predicts a most probable element sequence Y = (y1, y2, ..., y...) based on the input text X and instruction I. m ), each y t Let represent the t-th element to be predicted. Its probability can be expressed as: Among them, θ is the model parameter, y<t represents the first t-1 generated elements, and the model generates the most likely next element based on the input text and instructions. Represents the probability of generating the entire sequence Y given the input text X and instructions I, which is equal to the product of the probabilities of generating each element y in the sequence. The probability of the entire sequence Y is the product of the conditional probabilities of generating the next element y at each step. The model parameter θ determines the magnitudes of these probabilities. Through instruction fine-tuning, θ is adjusted to better generate the element sequence Y that conforms to the framework of the composite knowledge chain structure model according to the instructions I and the text X. t is the product of the probabilities. The probability of the entire sequence Y is the conditional probability of generating the next element y t at each step. The model parameter θ determines the magnitudes of these probabilities. Through instruction fine-tuning, θ is adjusted to better generate the element sequence Y that conforms to the framework of the composite knowledge chain structure model according to the instructions I and the text X.

[0042] Select a large language model pre-trained on a large amount of text data, such as the LLaMA series, GPT series models, or a model that has been further pre-trained or fine-tuned in the field of scientific literature to enhance its domain adaptability. The selected large language model has powerful natural language understanding capabilities and can deeply understand and generate text data in a specific domain. For example, use a GPT-3 model fine-tuned on scientific literature, which can understand and process specific terms and expressions in the scientific research field. To make the large language model more adaptable to the scientific research literature field, a model fine-tuned on computer science or biological literature can also be selected.

[0043] Design detailed and structured extraction instructions for the four core components of the problem component, method component, environment component, and result component and their important sub-elements (such as research objectives, key technologies, experimental conditions, findings, etc.). Use samples containing (instruction, literature fragment, expected extracted component element) to fine-tune the language large model. For example, extract the research objective sub-element of the problem component: From the following literature abstract and introduction, extract the specific scientific problems or technical objectives that the author clearly states this study aims to solve, and list each problem point or objective point as an independent entry. Extract the key technology sub-element of the method component: Identify and describe in detail the core research methods or key technical means proposed in this article. If there are multiple components or algorithms, list them separately and briefly describe their functions. Extract the experimental condition sub-element of the environment component: Find the text content that describes the experimental settings of this study, the details of the used dataset (such as name, scale, source), key parameter configurations, software tools, or specific geographical area background. Extract the finding sub-element of the result component: Summarize the main experimental results, numerical indicators, important findings, or core observed phenomena reported in this article. If specific performance data or comparisons are included, extract them together.

[0044] The preprocessed text units and corresponding instructions are input into the fine-tuned language model. The language model processes the documents according to the instruction set, extracts predefined component elements, and generates candidate component element texts. For example, for the text: To address the problem of low accuracy of existing urban expansion monitoring models in complex surface areas (the research objective of the problem component), this study proposes a dynamic monitoring method that integrates multi-source remote sensing data with deep convolutional neural networks (DCNN) (the key technology of the method component). The experiment was conducted in the Pearl River Delta region (the disciplinary background of the environmental component), using Landsat series imagery and nighttime light data (the tools used in the method component). The results show that this method improves the accuracy of urban built-up area extraction by 5.7% compared to traditional methods (the discovery of the result component, equivalent to the observation phenomenon). The ideal output of the large language model might be as follows: Research objective of the problem component: addressing the low accuracy of existing urban expansion monitoring models in complex terrain areas; Key technology of the method component: a dynamic monitoring method integrating multi-source remote sensing data and deep convolutional neural networks (DCNN); Disciplinary background of the environmental component: the Pearl River Delta region; Tools used in the method component: Landsat series imagery and nighttime light data; Findings of the results component: a 5.7% improvement in the accuracy of urban built-up area extraction.

[0045] The text fragments extracted by the language model undergo preliminary standardization, such as removing redundant expressions and standardizing terminology. To ensure the accuracy and consistency of terminology, a domain dictionary (such as a dictionary for geographic information science) can be used. The knowledge extraction result is structured data formed by the language model after processing scientific and technological literature, based on the instruction set and the composite knowledge chain structure model. This structured data corresponds to the various components and sub-elements in the composite knowledge chain structure model. For example, the target paper is input into the finely tuned language model. The language model scans the entire text, finding background information and research motivation in the introduction, precisely extracting specific values ​​such as temperature (250℃), pressure (5MPa), and flow rate (50mL / min) from the operational steps in the methods section, and extracting data such as 87%, 30%, and 120g / L / h from the experimental data in the results section. The final knowledge extraction result is a structured data block that clearly labels which component and sub-element of the composite knowledge chain structure model each extracted piece of information belongs to. For example: the research objective of the problem component is to increase the methanol selectivity to over 85%; the operation steps of the method component are: temperature 250℃, pressure 5MPa, and gas flow rate 50mL / min; and the experimental data of the result component are: methanol selectivity 87%, CO2 conversion rate 30%, and space-time yield 120g / L / h.

[0046] Through a clear set of instructions and targeted fine-tuning, the language big data model accurately identifies and distinguishes different types of information in documents, classifying them into the corresponding components and sub-elements of the composite knowledge chain structure model. This overcomes the problems of information omission, misclassification, or format confusion that may occur in traditional methods due to ambiguous text, scattered information, or lack of structured guidance. The generated structured knowledge extraction results are the direct raw materials for constructing a complete composite knowledge chain. Only by accurately and comprehensively extracting the information of each component and sub-element can the subsequent analysis of the internal relationships of components and the logical relationships across components be effectively carried out.

[0047] Based on the knowledge extraction results, the internal structural relationships and cross-component logical relationships of the components are parsed to generate the internal element relationships and cross-component logical relationships of the components.

[0048] Furthermore, this application also includes the following steps: parsing the hierarchy, temporal sequence, and dependency relationships between elements within the same component in the knowledge extraction results to generate the relationships between elements within the component; using preset prompts or instructions, calling a pre-trained classifier to determine whether there is a predefined logical relationship between two or more extracted component elements to generate the cross-component logical relationship.

[0049] Specifically, after identifying the knowledge extraction results, further analysis is performed on the hierarchical, temporal, or dependency relationships between elements within the same component. For example, specific instructions can be used to guide the sequence of steps within a language large-scale model identification method component, or NLP techniques such as dependency parsing can be combined. Dependency parsing is a natural language processing technique used to analyze the dependency relationships between words in a sentence. Revealing structural relationships in text through dependency parsing, such as which words depend on other words and their grammatical relationships, is extremely useful for analyzing the sequence of steps and dependencies in documents.

[0050] The internal element relationships of a component refer to the hierarchical structure, execution order, and dependencies among its various sub-elements (such as method steps) within the same core knowledge component (such as a method component). For example, in a method component, it may be necessary to first collect data, then perform preprocessing, and finally apply machine learning algorithms; there are clear sequential and dependency relationships between these steps.

[0051] Hierarchical relationships are the hierarchical relationships between sub-elements within the same component. For example, the steps in a method component may have a sequential order, with some steps depending on previous steps. Temporal relationships are the dependencies between sub-elements within the same component that occur in chronological order. For example, in a method component, some steps must be completed before subsequent steps can proceed. Dependencies are the mutual dependencies between sub-elements within the same component. For example, in a method component, a step may depend on the results of other steps to execute.

[0052] The language model is utilized again to guide the recognition of steps or components through specific instructions. For example, for the key technology of the extracted seemingly components: the dynamic monitoring method that integrates multi-source remote sensing data with deep convolutional neural networks (DCNN), the instruction is designed as follows: Please find and list the key processing flow or component modules of this method in the original text in sequence.

[0053] Once the hierarchical, temporal, and dependency relationships within a component are identified, these relationships are represented using graph structures (such as ordered lists, tree structures, or directed acyclic graphs (DAGs)) to clearly demonstrate the interrelationships between elements within the component. For example, the sequence of steps in a method component can be represented using a DAG: data preprocessing → feature extraction → model training, with arrows indicating dependencies from one step to the next. For instance, feature extraction depends on the results of data preprocessing, and model training depends on the results of feature extraction. After analyzing the internal structure of each component, a detailed graph of the relationships between internal elements is generated, clearly showing the hierarchy, temporal sequence, and dependencies between each component and its sub-elements.

[0054] Based on the knowledge extraction results (such as instances of problem components, method components, environment components, and result components), candidate relation pairs or tuples are formed by combining or combining predefined sets of logical relation types (such as drivers, constraints, and generation). For example, the research objectives of the problem component and the key technologies of the method component; the key technologies of the method component and the discovery of the result component; the experimental conditions of the environment component and the key technologies of the method component.

[0055] For each candidate relation pair, design specific prompts or instructions to guide the language model in determining whether a predefined logical relationship exists between them. For example, input prompt: Known problem: Optimize photovoltaic cell efficiency; Method: Deep Convolutional Neural Network (DCNN). Please determine whether this method is primarily used to solve this problem. Options: Yes / No / Uncertain. The language model determines whether a driving relationship or other logical relationship exists between the problem and the method by answering "Yes," "No," or "Uncertain." For example, known method: Deep Convolutional Neural Network (DCNN); Result: Efficiency improved by 10%. Please determine whether this method affects the result. Options: Yes / No / Uncertain. Alternatively, the language model can be directly guided to select the most suitable one from predefined relation types (such as driving relationship, solution relationship, etc.).

[0056] Candidate relationship pairs and their corresponding hints or instructions are input into a pre-trained classifier. The classifier determines whether a predefined logical relationship exists between two component elements based on the input and outputs the result. The candidate relationship pairs (C...) i C j Using these as inputs, train a classifier to predict the type of relationship between them, r∈R.inter These include driving relationships, constraint relationships, influence relationships, resolution relationships, derivation relationships, and support or challenge relationships. The model input can be an encoded representation of the component text, and the loss function typically uses cross-entropy loss. Where D is the training dataset, Φ is the parameter of the relation classification model, and r is the parameter of the model. true It is a real tag, that is, a component pair (C i C j The true relationship type between ).

[0057] Using a natural language generation model, a large language model is guided to generate sentences describing the relationships between components, and the relationship types are parsed from these sentences. For example, given the input prompt: Known problem: Optimize photovoltaic cell efficiency; Method: Deep Convolutional Neural Network (DCNN). Please generate a statement describing how this method solves this problem. The output of the large language model is: The method Deep Convolutional Neural Network (DCNN) was proposed to solve the problem of optimizing photovoltaic cell efficiency. From the generated sentences, the driving relationships are parsed and labeled as relationship types.

[0058] To improve the accuracy of relation classification, the language big data model is provided with contextual information about component elements, including the complete sentences or paragraphs within the original text. This allows the language big data model to understand the background and practical application of each element, thus making more accurate relation judgments. When judging the relationship between methods and results, the input text contains not only the components of the method and result but also their context in the literature. For example, given a description of the method Deep Convolutional Neural Network (DCNN) and the context of experimental results: applying the DCNN method under high-temperature conditions shows that the photovoltaic cell efficiency is improved by 10%. After understanding the complete context, the language big data model can determine that the DCNN method has an impact on the improvement of photovoltaic cell efficiency.

[0059] Based on the relation classification / generation results of the language big model, the final generated cross-component logical relationship can be represented as directed edges connecting components, generating cross-component logical relationships and comprehensively describing the interaction between various elements in scientific research activities.

[0060] The components, the relationships between elements within the components, and the logical relationships across components are structured and organized to generate a composite knowledge chain.

[0061] Further details are attached. Figure 2 As shown, this application also includes the following steps: defining logical rules to perform consistency verification and conflict resolution on the components, the relationships between elements within the components, and the cross-component logical relationships, and generating verification fusion results; constructing a component relationship graph based on the verification fusion results, using the identified component elements as nodes and the relationships as edges; assigning confidence scores to each element and relationship in the component relationship graph, and generating the composite knowledge chain.

[0062] Furthermore, this application also includes the following steps: constructing a multi-factor confidence assessment model; wherein the multi-factor confidence assessment model is constructed based on the output probability of a large language, the text support span, the matching correspondence with existing knowledge graphs, and the frequency of occurrence in multiple relevant documents; and assigning a confidence score to each element and relation in the component relationship graph using the multi-factor confidence assessment model.

[0063] Furthermore, this application also includes the following steps: the composite knowledge chain is stored in the form of a graph database, supporting the execution of scientific research problem analysis, result tracking, relationship chain mining and knowledge recommendation through a structured query language; wherein, the structured query language includes structural representations based on structural perspective, temporal perspective or topic perspective.

[0064] Specifically, the extracted components, their internal relationships, and cross-component logical relationships are structured and organized to form a complete and consistent composite knowledge chain. Logical rules (such as a method must be associated with a problem) are defined to verify the consistency and resolve conflicts of the extracted results from the large language model (i.e., components, their internal relationships, and cross-component logical relationships). These logical rules are a series of constraints set based on domain knowledge, research norms, or prior assumptions. They are used to verify the consistency of each component, its internal relationships, and cross-component logical relationships, ensuring that the relationships between components are reasonable and conform to the logic of actual scientific research activities when constructing the composite knowledge chain.

[0065] The purpose of consistency checks is to verify whether the extracted components and their relationships conform to predefined logical rules, ensuring that each component element and the relationships between them meet certain basic assumptions. For example, the relationship between method and problem: each method must be associated with a problem, that is, the method exists to solve a problem; the relationship between method and environment: each method is usually experimented with or applied in a specific environment; the relationship between result and method: each result is usually produced by a method in a specific environment.

[0066] When conflicts arise between different text fragments or extraction steps (such as inconsistent determinations of the relationship between the same pair of components), a confidence assessment mechanism can be used to select the result with higher confidence, or manual verification can be performed to resolve the conflict. For contradictory judgments that may arise in different stages of the language model or from different text fragments (such as inconsistent determinations of the relationship between the same pair of components), a confidence assessment mechanism can be introduced to select the judgment with higher confidence, or manual verification can be performed when necessary to resolve the conflict.

[0067] After consistency verification and conflict resolution, the final output is called the verification fusion result. It includes the confirmation status of each component element, the type of relationship between them, and eliminates any logical conflicts that may have existed before. All verified and conflict-resolved component elements and relationships are integrated into a structured knowledge graph, where each node represents a component (such as problem, method, environment, result), and each edge represents the logical relationship between them (such as driver, influence, constraint, etc.), thus obtaining the verification fusion result.

[0068] Based on the verification and fusion results, the identified component elements are used as nodes, and relationships are used as edges to construct a component relationship graph, which displays the structural relationships between various component elements. In the graph, each component or component element is represented as a node, and directed edges represent directional relationships between components, such as the driving relationship from problem to solution. Labeled edges are labels attached to each edge to indicate the relationship, such as impact or solution.

[0069] A confidence score is assigned based on the credibility of the extracted elements and relationships, representing the reliability of each element and relationship. For example, based on the probability generated by a large language model or external validation, when generating text, the large language model assigns a probability value to each generated word or component, representing the likelihood of that word or component appearing: Confidence(element) = f(P) LLM (element|text)).

[0070] Here, LLM stands for Large Language Model, and Confidence(element) is the confidence level of an element, i.e., the reliability of the large language model's extraction of that element. A higher confidence level indicates greater confidence in the model's extraction of that element. f() represents a function that maps the probabilities generated by the large language model to a confidence value. P LLM (element|text) represents the probability generated by the language model, indicating the probability of extracting an element given the text. In other words, when processing text, the model assigns the probability of each element appearing based on context and existing knowledge. An element can be any component of a complex knowledge chain (such as a problem, method, environment, or result) or a sub-element of a component (such as problem objectives, method steps, environmental assumptions, etc.). Text is the raw input text, which can be fragments extracted from scientific literature, research papers, reports, etc.

[0071] A multi-factor confidence assessment model is constructed based on the output probability of large language, text support span, matching correspondence with existing knowledge graphs, and frequency of occurrence in multiple relevant documents. Text support span refers to the length or range of supporting information related to a certain element or relationship in the text. Generally, the more abundant and detailed the supporting information in the text, the more credible the information. A larger support span means more evidence is provided to support the correctness of the element or relationship. A knowledge graph is a structured knowledge representation system where nodes represent concepts or entities, and edges represent relationships between them. In knowledge graph matching, the model compares elements or relationships extracted from documents with concepts in existing knowledge graphs to determine the degree of matching. A high matching degree indicates that the extraction result is likely accurate. Evidence frequency refers to the frequency with which a certain element or relationship appears in multiple relevant documents. If an element or relationship is mentioned and verified in multiple documents, it indicates greater universality and reliability, and therefore should be assigned a higher confidence level.

[0072] Construct a dataset containing multiple known elements and their true relationships. Each data point includes multiple input factors (such as the output probability of a large language model, text support span, etc.) and the true confidence score for that element. Train a confidence regressor using a regression algorithm (such as linear regression, support vector machine regression, etc.). The goal is to minimize the error between the predicted confidence score and the true confidence score. Typically, the loss function uses mean squared error to measure the difference between the regressor's predicted confidence score and the actual confidence score.

[0073] After training, the confidence regressor generates a confidence score for each feature or relation. By setting a threshold, only features and relations with high confidence scores are retained. For example, setting a confidence threshold (e.g., 0.85) will only retain relations and features with confidence scores higher than this value, filtering out the rest. If the same feature or relation is obtained from multiple documents or extraction steps, and their confidence scores differ, multiple results can be combined using a weighted fusion method. During fusion, weights can be applied based on confidence levels to ensure that results with higher confidence scores have a greater impact on the final knowledge chain.

[0074] According to the multi-factor confidence assessment model, each element and relation in the component relationship graph is assigned a confidence score. The confidence score is typically a numerical value between 0 and 1, representing the level of reliability. A higher confidence score indicates that the model has greater confidence in the reliability of the extraction result. First, the output probability generated by the large language model for each element (such as sub-elements of the problem component, method component, result component, and environment component) is calculated. The text support span is calculated. If the method description in the literature exceeds 300 words, indicating a large support span, the confidence score for that method will increase. The matching degree between the element or relation and the knowledge graph is calculated; evidence frequency is considered. By weighting and synthesizing these factors, a confidence score is finally generated for each element and relation. For example, the method component Deep Convolutional Neural Network (DCNN) received a comprehensive confidence score of 0.92, indicating that it is very reliable in the extraction result.

[0075] For example, assume the extracted components and relationships are as follows: Problem component P: improve energy storage efficiency; Method component M: deep learning algorithm; Result component R: efficiency improvement of 10%; Relationships: P→M, method drives problem, confidence level: 0.85, M→R, method affects result, confidence level: 0.92; Constructed composite knowledge chain graph: Nodes: P = improve energy storage efficiency, M = deep learning algorithm, R = efficiency improvement of 10%; Edges: P→M (driving relationship), M→R (affecting relationship); Confidence level of each edge: P→M (0.85), M→R (0.92).

[0076] Based on the above evaluation and assignment process, a composite knowledge chain is ultimately generated, supporting advanced knowledge services such as in-depth literature analysis, in-depth reasoning and question answering for research questions, and precise knowledge recommendation and discovery. The composite knowledge chain includes nodes (elements) and variables (relationships). Nodes comprise each extracted component element, and each node has an associated confidence score representing the reliability of that element. Edges represent the relationships between nodes, such as the driving relationship between a method and a problem, or the influence relationship between a method and a result. Each edge also has a confidence score representing the reliability of that relationship.

[0077] Precision, recall, and F1 score are used to evaluate component feature extraction. Precision = |TP| + |FP| |TP|, Recall = |TP| + |FN| |TP|, F1 = 2·Precision + Recall·Precision·Recall, where TP (True Positives) are correctly extracted features, FP (False Positives) are incorrectly extracted features, and FN (False Negatives) are unextracted features. Similarly, precision, recall, and F1 score are used to evaluate relation extraction, requiring determination of the relation type, direction, and whether the involved components are correct.

[0078] A graph-matching-based similarity metric is used to compare the composite knowledge chain with the gold standard, or the completeness, accuracy, and logical consistency of the composite knowledge chain are evaluated manually. The causes of extraction errors are analyzed (such as ambiguous instructions, misunderstandings in the language model, and non-standard original text expression), and based on this, targeted improvements are made to instruction design, optimization of language model fine-tuning data and strategies, or enhancement of post-processing rules, through iterative optimization to improve the overall quality of the constructed composite knowledge chain.

[0079] Storing complex knowledge chain data in a graph database enables efficient management and querying of knowledge elements and their relationships within research literature. Each element of a research activity (e.g., problem, method, result) is represented as a node in the graph, and each relationship (e.g., method drives problem, method influences result) is connected to the node as a labeled edge. The graph database facilitates complex graph traversal, pathfinding, and relationship mining operations. It supports various complex query operations, including research problem analysis, result tracking, relationship mining, and precise knowledge recommendation. The graph database provides specialized query languages ​​(e.g., Cypher, Neo4j's query language) to perform these operations.

[0080] When querying complex knowledge chains, structured representations can be taken from different perspectives to meet diverse research needs. While the structure of complex knowledge chains is suitable for knowledge reasoning and question answering, the required structure / representation may differ when used for research recommendation or research trend prediction. Structural perspective: complete component graph; temporal perspective: timeline-based research evolution chain; topical perspective: complex knowledge chain clusters clustered by question.

[0081] By storing composite knowledge chains in a graph database and supporting structured query languages ​​based on structural, temporal, or thematic perspectives, the depth and accuracy of scientific research logic representation are significantly improved. Through structured problem, method, environment, and outcome components—quadruple sets—along with their internal elements and interrelationships, the inherent logic of scientific research activities can be represented more meticulously and comprehensively. Based on instruction fine-tuning of a large language model and specific extraction strategies, composite knowledge chains can be automatically constructed from literature, improving efficiency.

[0082] In summary, the composite knowledge chain construction method for scientific research logical representation provided in this application has the following beneficial effects:

[0083] A composite knowledge chain structure model is constructed by defining core knowledge components. Based on this model, an instruction set for the components and sub-elements is built, and a pre-trained large language model is fine-tuned. The fine-tuned model then processes scientific and technological documents to generate knowledge extraction results. Based on these results, the internal structural relationships and cross-component logical relationships of the components are analyzed to generate internal element relationships and cross-component logical relationships. Finally, the components, their internal element relationships, and cross-component logical relationships are structured to generate a composite knowledge chain. In other words, by defining core knowledge components to construct a composite knowledge chain structure model and fine-tuning a pre-trained large language model to process scientific and technological documents, analyze the relationships between internal elements and cross-component logical relationships, and automatically construct a composite knowledge chain, the accuracy of knowledge extraction is improved.

[0084] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0085] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for constructing a composite knowledge chain oriented towards scientific research logical representation, characterized in that, include: A composite knowledge chain structure model is constructed by defining core knowledge components; Based on the aforementioned composite knowledge chain structure model, an instruction set for components and sub-elements is constructed, and the pre-trained large language model is fine-tuned using these instructions. The fine-tuned large language model is then used to process scientific and technological literature to generate knowledge extraction results. The components include a problem component, a method component, an environment component, and a result component. The sub-elements of the problem component include background information, research motivation, research gaps, and research objectives. The sub-elements of the method component include theoretical basis, key technologies, operational steps, and tools used. The sub-elements of the environment component include experimental conditions, underlying assumptions, disciplinary background, and constraints. The sub-elements of the result component include findings, experimental data, conclusions, evaluation, and discussion. Based on the knowledge extraction results, the internal structural relationships and cross-component logical relationships of the components are parsed to generate the internal element relationships and cross-component logical relationships of the components. The components, the relationships between elements within the components, and the logical relationships across components are organized in a structured manner to generate a composite knowledge chain; The construction of a composite knowledge chain structure model by defining core knowledge components includes: The core knowledge components are hierarchically and sequentially associated with their internal elements to construct a structural relationship diagram; A set of logical relationship types for connecting the core knowledge components is preset, including driving relationships, constraint relationships, influence relationships, resolution relationships, derivation relationships, and support or challenge relationships; By combining the structural relationship diagram and the set of logical relationship types, a formal representation of the composite knowledge chain is generated, thus producing the composite knowledge chain structure model. The step of structuring and organizing the components, the relationships between elements within the components, and the logical relationships across components to generate a composite knowledge chain includes: Define logical rules to perform consistency checks and conflict resolution on the components, the relationships between elements within the components, and the cross-component logical relationships, and generate a check fusion result; Based on the verification and fusion results, the identified component elements are used as nodes and the relationships are used as edges to construct a component relationship graph; Assign a confidence score to each element and relationship in the component relationship graph to generate the composite knowledge chain.

2. The method for constructing a composite knowledge chain oriented towards scientific research logical representation as described in claim 1, characterized in that, Building a composite knowledge chain structure model by defining core knowledge components also includes: A hierarchical system of relationship types is introduced, including basic relationships, domain-wide common research behaviors, and specific research paradigm relationships; By analyzing research literature samples from multiple fields using the aforementioned hierarchical system, an interdisciplinary relationship sub-library is constructed to supplement the set of logical relationship types.

3. The method for constructing a composite knowledge chain oriented towards scientific research logical representation as described in claim 1, characterized in that, The formal representation of a composite knowledge chain includes a structured object containing core knowledge components, their internal elements, and the relationships between components.

4. The method for constructing a composite knowledge chain oriented towards scientific research logical representation as described in claim 1, characterized in that, Based on the knowledge extraction results, the internal structural relationships and cross-component logical relationships of the components are parsed to generate internal element relationships and cross-component logical relationships, including: The hierarchical, temporal, and dependency relationships among elements within the same component in the knowledge extraction results are analyzed to generate relationships among elements within the component. Using preset prompts or instructions, a pre-trained classifier is invoked to determine whether there is a predefined logical relationship between two or more extracted component elements, thereby generating the cross-component logical relationship.

5. The method for constructing a composite knowledge chain oriented towards scientific research logical representation as described in claim 1, characterized in that, Assign a confidence score to each element and relationship in the component relationship graph, including: Construct a multi-factor confidence assessment model; The multi-factor confidence assessment model is constructed based on the output probability of a large language, the text support span, the matching correspondence with existing knowledge graphs, and the frequency of occurrence in multiple relevant documents. The confidence score is assigned to each element and relationship in the component relationship diagram using the multi-factor confidence assessment model.

6. The method for constructing a composite knowledge chain oriented towards scientific research logical representation as described in claim 1, characterized in that, After generating the composite knowledge chain, it also includes: The composite knowledge chain is stored in the form of a graph database, which supports the analysis of scientific research questions, tracking of results, mining of relationship chains and knowledge recommendation through structured query language; Among them, the structured query language includes structured representations based on structural perspectives, temporal perspectives, or topical perspectives.

Images