Green carbon reduction key technology knowledge graph construction method, system, medium and equipment

Abstract

The invention discloses a green carbon-reduction key technology knowledge graph construction method and a system with the green carbon-reduction key technology knowledge graph construction method, the method is innovated and improved on the basis of a traditional Graph RAG method, and through introduction of Prompt Tuning and CoT mechanisms, the construction of the green carbon-reduction key technology knowledge graph is realized, so that the construction of the green carbon-reduction key technology knowledge graph is realized, and the construction efficiency of the green carbon-reduction key technology knowledge graph is improved. And the LLM is combined to realize automatic extraction and integration of green carbon reduction field knowledge. The constructed knowledge graph systematically integrates a core knowledge system, a technology development context, a subject evolution trend and a research frontier dynamic state in the field. Meanwhile, an intelligent question and answer system is constructed based on the green carbon reduction key technology knowledge graph, a CoT mechanism and a voting mechanism are coordinated, and the accuracy and comprehensiveness of knowledge retrieval and the availability of the knowledge graph are effectively improved. The method can provide effective decision support for technical development status carding and future trend analysis in the green carbon reduction field, and assists scientific research innovation and industrial development in the field.

Description

Technical Field

[0001] This invention relates to the field of energy information processing technology, and in particular to a method, system, medium, and equipment for constructing a knowledge graph of key green carbon reduction technologies. Background Technology

[0002] Green carbon reduction technologies, as a key means of addressing climate change, encompass several crucial areas, including carbon capture, utilization and storage (CCUS), renewable energy development and utilization, smart grid technology, and energy management. However, due to the multidisciplinary nature and diverse application scenarios involved in these technologies, related research findings are fragmented and heterogeneous, lacking a unified knowledge integration mechanism. This makes it difficult to systematically summarize, efficiently search, and fully utilize technical information.

[0003] Knowledge graphs, as an advanced knowledge organization and representation technology, enable systematic cognition and knowledge management in complex technological fields by constructing structured networks of entities, relationships, and attributes. Currently, knowledge graph construction primarily relies on manual annotation, natural language processing (NLP), and deep learning techniques. Existing knowledge graph construction methods mainly fall into the following categories:

[0004] (1) Rule-based method: This method relies on domain experts to formulate rules for entity extraction and relationship identification. It is suitable for small-scale data processing, but it has problems such as poor scalability and high rule formulation cost, making it difficult to adapt to the needs of large-scale knowledge graph construction.

[0005] (2) Machine learning-based methods: This method uses supervised or semi-supervised learning algorithms for entity recognition and relation extraction. Although it improves the level of automation to a certain extent, its performance is highly dependent on a large amount of labeled data and its generalization ability is limited, making it difficult to cope with the dynamic updates of domain knowledge.

[0006] (3) Methods based on pre-trained language models: This method uses large-scale pre-trained language models (such as BERT, GPT, etc.) for text understanding and relational reasoning, which significantly improves the automation efficiency of knowledge extraction. However, this method still has problems such as knowledge update lag, opaque reasoning chain, and insufficient domain adaptability, which limits its application effect in specific fields such as green carbon reduction.

[0007] In the application of knowledge graphs, existing technologies combine knowledge graphs with Large Language Models (LLMs) to form retrieval enhancement architectures, exemplified by Graph RAGs. As a structured resource repository of factual information, knowledge graphs are transformed into query information by LLMs, which retrieve relevant knowledge from the graph and generate answers to questions. Applying knowledge graphs to intelligent question-answering systems can support decision analysis and knowledge retrieval in the field of green carbon reduction. Summary of the Invention

[0008] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a knowledge graph construction and intelligent question-answering method for the field of green carbon reduction. It innovatively improves upon the traditional Graph RAG method by introducing prompt tuning and chain-of-thought (CoT) mechanisms, combined with a large language model (LLM), to achieve automated extraction and integration of knowledge in the field of green carbon reduction. The constructed knowledge graph systematically integrates the core knowledge system, technological development trajectory, disciplinary evolution trends, and cutting-edge research dynamics of this field. Simultaneously, an intelligent question-answering system is constructed based on the knowledge graph of key green carbon reduction technologies, in conjunction with the CoT and voting mechanisms, effectively improving the accuracy and comprehensiveness of knowledge retrieval and the usability of the knowledge graph. This invention can provide effective decision support for reviewing the current state of technological development and analyzing future trends in the field of green carbon reduction, thus contributing to scientific research innovation and industrial development in this field.

[0009] This invention also proposes a system with the aforementioned knowledge graph construction and intelligent question answering methods for the field of green carbon reduction.

[0010] The method for constructing a knowledge graph of key green carbon reduction technologies according to a first aspect of the present invention is characterized by comprising the following steps:

[0011] Acquire multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data;

[0012] The entity relationship extraction task is performed by driving the large language model through domain-adapted prompt templates, and the entity recognition and relationship extraction are optimized by combining the thinking chain technology. Knowledge graph elements and element summaries are generated based on the standardized resource base.

[0013] The knowledge graph and element summary are hierarchically divided into communities based on the community detection algorithm. A structured summary is generated for each community module, and a multi-granularity information mapping is constructed between the source document data layer, the knowledge graph layer and the community summary layer.

[0014] Based on the thought chain-based prompting engineering framework and hierarchical graph index architecture, and using the structured summary of the community module, a comprehensive answer integrating domain knowledge is generated using a large language model, and the optimal global answer is selected through a multi-round voting mechanism.

[0015] The method for constructing a knowledge graph of key green carbon reduction technologies according to embodiments of the present invention has at least the following beneficial effects: The present invention innovatively improves upon the traditional Graph RAG method by introducing prompt tuning and chain-of-thought (CoT) mechanisms, combined with a large language model, to achieve automated extraction and integration of knowledge in the field of green carbon reduction. The constructed knowledge graph systematically integrates the core knowledge system, technological development trajectory, disciplinary evolution trends, and cutting-edge research dynamics of this field. Simultaneously, an intelligent question-answering system is constructed based on the knowledge graph of key green carbon reduction technologies, coordinating the CoT and voting mechanisms to effectively improve the accuracy and comprehensiveness of knowledge retrieval and the usability of the knowledge graph. The present invention can provide effective decision support for reviewing the current state of technological development and analyzing future trends in the field of green carbon reduction, thus contributing to scientific research innovation and industrial development in this field.

[0016] According to some embodiments of the present invention, the data cleaning step in the step of acquiring multi-source heterogeneous data and cleaning and structuring the multi-source heterogeneous data to establish a standardized resource library includes:

[0017] Text extraction and noise removal are performed on HTML and PDF documents in the multi-source heterogeneous data.

[0018] Long documents are segmented using a text chunking strategy, with each text chunk having a capacity of 600 tokens and an overlap area of ​​100 tokens between chunks.

[0019] According to some embodiments of the present invention, the step of generating knowledge graph elements and element summaries by driving a large language model to perform entity relation extraction tasks through domain-adapted prompt templates, and optimizing entity recognition and relation extraction by combining mind chain technology, specifically includes the following steps:

[0020] Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators;

[0021] The prompt templates containing multi-step reasoning are designed using the thinking chain technology, combined with a small number of examples to guide a large language model, and entity relationships are extracted based on a well-established standardized resource library;

[0022] The extracted entity, relation, and covariate information are summarized by elements to construct a knowledge graph in the form of an isomorphic undirected weighted graph.

[0023] According to some embodiments of the present invention, in the step of performing hierarchical community division of the knowledge graph and element summary based on the community detection algorithm, generating a structured summary for each community module, and constructing a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer, the process of generating a structured summary for each community module and constructing a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer specifically includes:

[0024] In the Leaf Layer community, element digests are added in ascending order of node degree priority until the token limit is reached;

[0025] In a high-level community, if the total length of all element summaries within the community does not exceed the token limit, then the element summaries within the high-level community are added in the original order. If the total length of all element summaries within the community exceeds the token limit, then the sub-community element summaries are sorted from longest to shortest, and the shorter sub-community summaries replace the corresponding longer element summaries, gradually reducing the number of tokens until the token limit is met.

[0026] According to some embodiments of the present invention, in the steps of generating a comprehensive answer integrating domain knowledge using a large language model based on the structured summary of the community module and the hierarchical graph index architecture based on the thought chain-based prompting engineering framework and the hierarchical graph index architecture, and selecting the optimal answer through a multi-round voting mechanism, the process of selecting the optimal global answer through a multi-round voting mechanism specifically includes:

[0027] The structured summary of the community is randomly shuffled and divided into several community summary blocks;

[0028] Intermediate answers are summarized in descending order of scores. The LLM is forced to follow the reasoning process of "problem analysis → information filtering → content integration → logical arrangement" by combining the thinking chain framework and generating intermediate answers and usefulness scores based on the summary blocks of each community.

[0029] Filter out intermediate answers with a score of 0, and generate a global answer by sorting the intermediate answers from each community in descending order of their usefulness scores, until the token limit is reached.

[0030] The intelligent question-and-answer system for green carbon reduction key technologies knowledge graph according to a second aspect of the present invention is characterized by comprising:

[0031] The data processing module is used to acquire multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data.

[0032] The knowledge graph construction module is used to drive the large language model to perform entity relation extraction tasks through domain-adapted prompt templates, optimize entity recognition and relation extraction by combining thinking chain technology, and generate knowledge graph elements and element summaries based on the standardized resource library.

[0033] The community detection module can perform hierarchical community division of the knowledge graph and element summary based on the community detection algorithm, generate a structured summary for each community module, and construct a multi-granularity information mapping between the source document data layer, the knowledge graph layer and the community summary layer.

[0034] The question-answering implementation module is able to generate comprehensive answers that integrate domain knowledge based on the prompting engineering framework of the thinking chain and the hierarchical graph index architecture, and based on the structured summary of the community module, using a large language model, and select the optimal global answer through a multi-round voting mechanism.

[0035] Furthermore, in the data processing module, data cleaning includes:

[0036] Text extraction and noise removal are performed on HTML and PDF documents in the multi-source heterogeneous data.

[0037] Long documents are segmented using a text chunking strategy, with each text chunk having a capacity of 600 tokens and an overlap area of ​​100 tokens between chunks.

[0038] Furthermore, in the knowledge graph construction module, the process of generating knowledge graph elements and element summaries specifically includes:

[0039] Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators;

[0040] The prompt templates containing multi-step reasoning are designed using the thinking chain technology, combined with a small number of examples to guide a large language model, and entity relationships are extracted based on a well-established standardized resource library;

[0041] The extracted entity, relation, and covariate information are summarized by elements to construct a knowledge graph in the form of an isomorphic undirected weighted graph.

[0042] Furthermore, in the community detection module, the process of generating a structured summary for each community module and constructing a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer specifically includes:

[0043] In the Leaf Layer community, element digests are added in ascending order of node degree priority until the token limit is reached;

[0044] In a high-level community, if the total length of all element summaries within the community does not exceed the token limit, then the element summaries within the high-level community are added in the original order. If the total length of all element summaries within the community exceeds the token limit, then the sub-community element summaries are sorted from longest to shortest, and the shorter sub-community summaries replace the corresponding longer element summaries, gradually reducing the number of tokens until the token limit is met.

[0045] Furthermore, in the question-answering implementation module, the process of selecting the optimal answer through a multi-round voting mechanism specifically includes:

[0046] The structured summary of the community is randomly shuffled and divided into several community summary blocks;

[0047] Intermediate answers are summarized in descending order of scores. The LLM is forced to follow the reasoning process of "problem analysis → information filtering → content integration → logical arrangement" by combining the thinking chain framework and generating intermediate answers and usefulness scores based on the summary blocks of each community.

[0048] Filter out intermediate answers with a score of 0, and generate a global answer by sorting the intermediate answers from each community in descending order of their usefulness scores, until the token limit is reached.

[0049] According to a third aspect of the present invention, a computer-readable storage medium stores computer-executable instructions for performing the above-described method for constructing a knowledge graph of key green carbon reduction technologies.

[0050] According to a fourth aspect of the present invention, the terminal includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for constructing a knowledge graph of key green carbon reduction technologies.

[0051] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0052] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0053] Figure 1 This is a schematic diagram illustrating the steps of the method for constructing a knowledge graph of key green carbon reduction technologies according to an embodiment of the present invention.

[0054] Figure 2 A detailed technical roadmap for the method of constructing a knowledge graph of key green carbon reduction technologies provided in the embodiments of the present invention;

[0055] Figure 3This invention provides a green carbon reduction knowledge graph in the method for constructing a knowledge graph of key green carbon reduction technologies.

[0056] Figure 4 The structural block diagram of the green carbon reduction key technology knowledge graph intelligent question answering system provided in the embodiments of the present invention. Detailed Implementation

[0057] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0058] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0059] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0060] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0061] Example 1

[0062] Existing knowledge graphs in the field of green carbon reduction suffer from incomplete knowledge coverage and a lack of systematic knowledge framework, making it difficult to meet the needs of interdisciplinary technical analysis and application. Furthermore, traditional knowledge graph construction methods have limitations in terms of domain adaptability, particularly the poor performance of existing Graph RAG methods in key entity identification and extraction in the green carbon reduction field. Meanwhile, traditional question-answering systems built on traditional knowledge graphs suffer from technical bottlenecks due to their reliance on keyword matching or traditional RAG methods, resulting in insufficient accuracy and incomplete knowledge coverage.

[0063] To address the problems of existing technologies, this application provides a method for constructing a knowledge graph of key green carbon reduction technologies, such as... Figure 1 As shown, it includes at least:

[0064] Step S100: Obtain multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data.

[0065] By collecting heterogeneous data from multiple sources in the field of green carbon reduction, including but not limited to scientific literature, policies and regulations, patent databases and industry dynamic data, a standardized resource library is established after data cleaning and structuring.

[0066] Step S200: Drive the large language model to perform entity relation extraction task through domain-adapted prompt templates, optimize entity recognition and relation extraction by combining thinking chain technology, and generate knowledge graph elements and element summaries based on the standardized resource base.

[0067] By introducing the MindChain technology, we can further construct domain-adaptive prompt templates to drive the LLM to perform entity relationship extraction tasks, thereby building a knowledge graph.

[0068] Step S300: Based on the community detection algorithm, perform hierarchical community division on the knowledge graph and element summary, generate a structured summary for each community module, and construct a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer.

[0069] A hierarchical graph index architecture is established. Based on a community detection algorithm, the knowledge graph is divided into communities. For each community module, a structured summary is generated using LLM based on the entities, relationships, and their descriptions within the community. A bidirectionally linked graph structure index is then constructed, thereby achieving multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer.

[0070] Step S400: Based on the thought chain-based prompting engineering framework and hierarchical graph index architecture, and based on the structured summary of the community module, a comprehensive answer integrating domain knowledge is generated using a large language model, and the optimal global answer is selected through a multi-round voting mechanism.

[0071] We design a prompting engineering framework based on thought chains, combining it with an existing graph index architecture. Leveraging the generative capabilities of LLM, we output comprehensive answers that integrate domain knowledge, policy norms, and technical pathways, and provide interpretable traceability of the decision-making basis. Furthermore, by integrating a voting mechanism, we further improve the quality of responses.

[0072] Based on the above embodiments, this application innovatively improves upon the traditional Graph RAG method by introducing prompt tuning and chain-of-thought (CoT) mechanisms, combined with LLM, to achieve automated extraction and integration of knowledge in the field of green carbon reduction. The constructed knowledge graph systematically integrates the core knowledge system, technological development trajectory, disciplinary evolution trends, and cutting-edge research dynamics of this field. Simultaneously, an intelligent question-answering system is constructed based on the knowledge graph of key green carbon reduction technologies, in conjunction with the CoT and voting mechanisms, effectively improving the accuracy and comprehensiveness of knowledge retrieval and the usability of the knowledge graph. This invention can provide effective decision support for reviewing the current state of technological development and analyzing future trends in the field of green carbon reduction, thus contributing to scientific research innovation and industrial development in this field.

[0073] Example 2

[0074] Furthermore, in order to better demonstrate the purpose of this application, a more detailed description will be provided based on Embodiment 1.

[0075] like Figure 2 As shown, embodiments of this application include the following processes:

[0076] Step S100: Obtain multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data.

[0077] For key green carbon reduction technology documents in various formats such as HTML and PDF obtained from the internet, data quality optimization is performed to address unstructured and noisy issues. This includes data extraction and data cleaning. Data extraction primarily involves using methods such as text extraction and webpage parsing to obtain key text from the source documents. Data cleaning mainly involves noise removal and stop word filtering to eliminate irrelevant text elements such as special abbreviations and special symbols.

[0078] Step S200: Drive the large language model to perform entity relation extraction task through domain-adapted prompt templates, optimize entity recognition and relation extraction by combining thinking chain technology, and generate knowledge graph elements and element summaries based on the standardized resource base.

[0079] To resolve the conflict between the LLM context window limitation and long document processing during subsequent entity relation extraction, the source document is segmented into text blocks to determine a reasonable text segmentation granularity that balances information integrity and computational efficiency. Specifically, the Python nltk library is used for text segmentation, and the segmentation strategy is as follows:

[0080] (1) Block size control: The capacity of a single text block is set to 600 tokens, and a sentence-by-sentence accumulation strategy is adopted to ensure that each text block reaches the preset size;

[0081] (2) Semantic continuity guarantee: Set up an inter-block overlap area of ​​100 tokens to maintain the semantic relevance across text blocks.

[0082] Step S300: Based on the community detection algorithm, perform hierarchical community division on the knowledge graph and element summary, generate a structured summary for each community module, and construct a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer.

[0083] The resource repository is manually analyzed to identify entity types used to describe the field of green carbon reduction. Next, a thought-chain mechanism is used for prompt optimization, constructing a domain-adaptive entity relationship prompting engineering. Then, the segmented source text blocks are sequentially input into an LLM (Local Level Management) system for entity and relationship extraction, finally converting them into structured knowledge graph elements and generating element summaries. Specifically, this includes:

[0084] Step S301: Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators.

[0085] Based on thematic and content analysis of resource databases related to green carbon reduction, highly relevant terms to green carbon reduction technologies are extracted from the text and categorized to construct a dedicated entity type list, Entity_types, applicable to this field. The constructed list includes: [technology, system, organization, substance / material, policies / regulations, project, equipment, effect / metric, energy_source]. The optimized entity type list adds key categories such as technology, policy, and material, covering key knowledge elements in the field of green carbon reduction, including technological systems, energy sources, and performance indicators.

[0086] Step S302: Design a prompt template containing multi-step reasoning using the thinking chain technology, and combine it with a small number of examples to guide the large language model to extract entity relationships.

[0087] To help the model better understand the task requirements of text and entity relation extraction in the green hydrogen field, this invention combines thought chain technology for prompt optimization and reconstructs the prompt template for entity relation extraction, effectively improving the entity recognition accuracy of LLM in the green carbon reduction field. The specific technical solution is as follows:

[0088] (1) Prior Knowledge Injection. To enable the LLM to more accurately understand each entity type, the newly defined green carbon reduction entity types are described in multiple dimensions in the prompts, including but not limited to: entity terminology definitions, functional attribute descriptions, and typical instance enumerations. By adding a complete descriptive framework for entity types, a domain knowledge representation benchmark is established, improving the entity recognition performance of large models. For example, the description of the entity type substance / material is as follows:

[0089] substance / material: A chemical element, compound, or engineered material that plays a critical role in CCUS and green energy applications. This includes gases, solvents, catalysts, adsorbents, and structural materials used in carbon capture, storage, and utilization processes. Examples: CO2, H2, CH4, Aminesolvents, Metal-organic frameworks (MOFs), Catalysts for hydrogen production.

[0090] (2) Developing a thought-reasoning framework. To improve the interpretability of the LLM reasoning process and the accuracy of entity recognition, this invention introduces thought chain technology and designs a thought-reasoning framework containing eight reasoning steps, decomposing the entity recognition task into an interpretable reasoning process. The thought-reasoning framework is as follows:

[0091] First, it suggests that LLM should perform semantic filtering based on the green carbon reduction domain to reduce misidentification of general domain entities and irrelevant entities. Second, it suggests that LLM should implement a three-level entity quality assurance system consisting of text source verification, type mapping verification, and decoupling of coupled entities. Finally, it suggests that LLM should perform output format verification to ensure that the generated knowledge extraction results strictly conform to standard formats.

[0092] (3) Few-shot guidance for model understanding of tasks. To enhance the LLM's ability to recognize text in the field of green carbon reduction and optimize its entity relationship extraction performance based on thought chain, this invention provides four sets of standardized input and output examples for the field of green carbon reduction in the prompt template.

[0093] Specifically, each set of examples includes: original domain text fragments and a list of entity types, a thought chain reasoning process, and standard entity relationship recognition results. The four sets of examples provided cover all preset entity types and establish a standardized paradigm for knowledge extraction in the field of green carbon reduction through a complete mapping between input text and corresponding thought chain reasoning processes.

[0094] Step S303: Perform element summarization on the extracted entity, relation and covariate information, and construct a knowledge graph in the form of an isomorphic undirected weighted graph.

[0095] Based on the constructed domain knowledge extraction prompt template, LLM is used to identify all entities in the text, including their names, types, and descriptions. Then, all relationships between related entities are identified, including source and target entities and descriptions of their relationships. In addition to entities and relationships, other relevant information (covariates) related to node instances is extracted, which helps the model to understand the extracted node instances more comprehensively. Default covariates include subject, object, type, description, source text span, and its start and end dates.

[0096] Next, an element summary is performed on each extracted graph element (entity node, relation edge, and declared covariate), transforming instance-level information into a more concise and descriptive text block. Utilizing rich descriptive text to represent nodes in the graph enables the model to better handle complex, potentially noisy data, while retaining the reasoning capabilities of traditional knowledge graphs.

[0097] The knowledge graph can be constructed as an isomorphic undirected weighted graph using the element indexes created earlier. Entity nodes are connected by relation edges, and the weight of each edge represents the normalized count of detected relation instances, reflecting the strength or frequency of the relation.

[0098] Step S304, Community Detection and Community Summary

[0099] After constructing a knowledge graph of key green carbon reduction technologies, community detection is used to identify clusters in different technical fields or research directions. The community structure of the knowledge graph enhances its interpretability and organization, facilitating subsequent searches and queries. A community is defined as a subset of nodes, with tightly connected nodes within a community exhibiting strong correlations, while connections to nodes in other communities are weaker.

[0100] This invention employs the Leiden algorithm to detect communities in a knowledge graph, dividing it into a hierarchical community structure. The Leiden algorithm gradually aggregates nodes into sub-communities through local optimization and iterative processes. Its main steps include: (1) local node movement; (2) partition refinement; and (3) network aggregation based on the refinement results. Through multiple iterations, a hierarchical community structure is finally obtained, with each level corresponding to a community partition, fully covering all nodes in the graph, and these partitions intersect and exhaustively cover each other.

[0101] The leaf-layer communities contain the finest-grained information, while the upper-layer communities provide a more abstract and comprehensive view. Subsequently, a bottom-up approach is used to generate a summary report for each community, with the following steps:

[0102] (1) Leaf layer community summary: Based on the priority of nodes, edges and covariates (priority is sorted according to the degree of nodes), add element summaries to the context window of LLM incrementally until the token limit is reached.

[0103] (2) High-level community summary: If the total length of all element summaries in the community does not exceed the token limit, the element summaries in the high-level community are added directly in order; if the limit is exceeded, the sub-community element summaries are sorted from longest to shortest according to their length, and the shorter sub-community summaries are used to replace the corresponding longer element summaries, gradually reducing the number of tokens until the token limit is met.

[0104] The community summary provides an overview of the entire dataset, helping the model understand the global structure and semantics of the dataset and providing useful contextual information for subsequent queries.

[0105] Step S400: Based on the thought chain-based prompting engineering framework and hierarchical graph index architecture, and based on the structured summary of the community module, a comprehensive answer integrating domain knowledge is generated using a large language model, and the optimal global answer is selected through a multi-round voting mechanism.

[0106] Step S410: Implementation of knowledge graph question answering.

[0107] When a user submits a query, for a given community level, a final global answer can be generated using the MapReduce mechanism. The MapReduce process is as follows:

[0108] Step S411, Random Shuffling and Chunking: First, the community summaries at this level are randomly shuffled to ensure a uniform distribution of information. Then, these summaries are divided into blocks of a predetermined size, thus distributing relevant information across multiple blocks instead of concentrating it in a single context window, which avoids information loss.

[0109] Step S412: Community generates intermediate answers: Each community summary block generates intermediate answers independently and in parallel using LLM. Simultaneously, LLM generates a usefulness score of 0-100 for each block's intermediate answer, indicating the degree of usefulness of the intermediate answer to the target question.

[0110] Step S413, Global Answer Summary: After filtering out answers with a score of 0, sort the intermediate answers from each community in descending order of usefulness score and add them step by step to a new context window until the token limit is reached. Once all relevant intermediate answers have been considered, generate the global answer to be returned to the user through the final context window.

[0111] To address the random volatility inherent in generative model outputs, this invention improves answer quality by integrating a thought chain reasoning framework with a multi-round voting mechanism. The specific solution is as follows:

[0112] First, the MindChain framework provides LLM with reasoning logic. Through a phased, structured reasoning process, it forces the model to follow a causal chain of "problem analysis → information filtering → content integration → logical arrangement," significantly reducing the risk of random output. Second, the MindChain framework instructs LLM to select data sources based on descending order of report importance. Through data citation integrity checks and modal verb retention strategies (such as "should," "may," and "will"), it ensures that all generated content is supported by traceable knowledge graph nodes, improving the credibility of the answers. Finally, it prompts LLM to organize responses logically, dividing them into several clear parts: introduction (background), main analysis (key insights), and impact and conclusion (importance), thereby ensuring coherent information and clear causal relationships in the answers.

[0113] Furthermore, combining the thought chain with a multi-round voting mechanism can generate multiple candidate answers that follow the same reasoning path. The optimal solution can be selected through voting, thereby improving the stability of the LLM-generated answers.

[0114] Step S420: Knowledge graph construction and question-answering evaluation metric construction. Specifically, this includes:

[0115] Step S421, Entity Recognition Evaluation Indicators.

[0116] The F1 score was used as the primary evaluation metric, and it was calculated based on the model entity extraction results and manual annotation results. The formula for calculating the F1 score is as follows:

[0117]

[0118] Where precision is the accuracy rate, reflecting the correctness of entity type matching; recall is the comprehensiveness of entity recognition. Precision and recall are calculated as follows:

[0119]

[0120] In the above formulas, TP indicates that the entity name is correctly identified and the type is correctly matched; FP indicates that the entity name is correctly identified but the type is incorrect; and FN indicates that the entity is not identified.

[0121] Step S422: Quality assessment indicators for answer generation.

[0122] The LLM scorer was used to assess the quality of the generated answers, evaluating the quality across the following four dimensions:

[0123] (1) Completeness: Whether all information in the question is parsed and a targeted answer is provided.

[0124] (2) Logicality: Whether the argument is rigorous and whether the conclusion is reasonable.

[0125] (3) Readability: Whether the answer structure is clear and whether the sentences are fluent and unambiguous.

[0126] (4) Accuracy (accuracy_sources): scored based on the number of cited reports.

[0127] The specific evaluation method is described in the following formula:

[0128]

[0129] Among them, Score i This represents the quality score of the answer to the i-th question. For the same question, different answers are input into the LLM for evaluation. The LLM will score the quality of the answer based on the four dimensions mentioned above, and the answer with the highest average score will be the excellent answer. At the same time, the excellence rate will be calculated.

[0130]

[0131] Among them, Score i,post Score represents the quality score of the improved answer to the i-th question. i,pre This represents the quality score of the answer to the i-th question before improvement.

[0132] Example 3

[0133] To verify the correctness of the above embodiments one and two, the present invention was verified through experiments. The present invention has achieved significant optimization in terms of entity recognition and question answering system quality, and the specific effects are as follows.

[0134] (1) Significantly improved accuracy in key entity recognition and enhanced domain adaptability:

[0135] To address the issue of unsatisfactory entity recognition performance of traditional methods in the field of green carbon reduction, this invention significantly improves the accuracy and recall rate of entity type matching by combining domain-optimized Prompt Tuning technology and a thought chain mechanism. Experimental data are shown in Table 1.

[0136] Table 1 Experimental Results of Improved Entity Recognition

[0137]

[0138] Experimental results show that the improved entity recognition precision increased by 7.34%, indicating a significant reduction in the entity type discrimination error rate; recall increased by 6.31%, proving that the system can recognize more correct entities; and the F1 score increased by 6.79%, comprehensively verifying the domain adaptability of the entity recognition model.

[0139] Experimental comparison results show that the above improvements enhance the knowledge extraction and integration capabilities, effectively solve the problem of incomplete knowledge coverage caused by entity omissions or type errors in knowledge graph construction, make the construction of knowledge graphs more complete, and provide a more systematic and complete data foundation for interdisciplinary technical analysis.

[0140] (2) The search quality of the intelligent question-answering system has been comprehensively improved:

[0141] To address the shortcomings in accuracy and coverage of traditional question-answering systems that rely on keyword matching or traditional RAG methods, this invention achieves a comprehensive improvement in answer generation quality by integrating the CoT mechanism and a voting mechanism. Experimental comparisons are shown in Table 2.

[0142] Table 2. Experimental Results of Question-Answer System Improvement

[0143]

[0144] Experimental results show that the average response quality improved by 2.98%, mainly in terms of completeness and accuracy, with improvements of 6.10% and 4.08% respectively. In terms of logic and readability, these two aspects are mainly limited by the LLM model itself. Therefore, it is difficult to improve the response quality in these two aspects through thought chain and voting mechanisms. Table 3 illustrates the comparison of response quality before and after the improvement.

[0145] Table 3. Examples of Q&A Performance Comparison

[0146]

[0147]

[0148]

[0149] (3) Completed the systematic construction of a knowledge graph of key technologies in the field of green carbon reduction, which has high practical application value:

[0150] Based on improved automated knowledge extraction technology, the knowledge graph constructed in this invention (see appendix) Figure 3 It can dynamically integrate the core technology system, discipline evolution trends and cutting-edge developments in the field of green carbon reduction, which can help scientific research and industrial applications. It can be used to analyze the current status of technological development and predict future trends in the field of green carbon reduction, and provide precise decision support for researchers and enterprises.

[0151] In summary, through technological innovation and experimental verification, this invention has achieved breakthrough improvements in core aspects such as entity recognition, knowledge integration, and question-answer generation, significantly improving the efficiency of knowledge acquisition and application, and constructing a systematic knowledge graph of key technologies in the field of green carbon reduction, providing strong support for the development of green carbon reduction technologies.

[0152] Another aspect of this application provides a green carbon reduction key technology knowledge graph intelligent question-answering system, such as... Figure 4 As shown, the system 40 includes:

[0153] Data processing module 401 is used to acquire multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data.

[0154] The knowledge graph construction module 402 is used to drive the large language model to perform entity relationship extraction tasks through domain-adapted prompt templates, and combine the thinking chain technology to optimize entity recognition and relationship extraction, and generate knowledge graph elements and element summaries.

[0155] The community detection module 403 is capable of hierarchically dividing the knowledge graph into communities based on the community detection algorithm, generating a structured summary for each community module, and constructing a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer.

[0156] The question-answering implementation module 404 is able to generate comprehensive answers that integrate domain knowledge based on the prompting engineering framework of the thinking chain and the hierarchical graph index architecture, and based on the standardized resource base, knowledge graph elements, element summaries, and structured summaries of the community module, using a large language model, and select the optimal answer through a multi-round voting mechanism.

[0157] Furthermore, the process of generating knowledge graph elements and element summaries in knowledge graph construction module 402 includes:

[0158] Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators;

[0159] The prompt template, which incorporates multi-step reasoning, is designed using the thinking chain technology, and combined with a small number of examples to guide the large language model in extracting entity relationships.

[0160] The extracted entity, relation, and covariate information are summarized by elements to construct a knowledge graph in the form of an isomorphic undirected weighted graph.

[0161] Furthermore, the data cleaning process in data processing module 401 includes:

[0162] Perform text extraction and noise removal on HTML and PDF documents;

[0163] Long documents are segmented using a text chunking strategy, with each text chunk having a capacity of 600 tokens and an overlap area of ​​100 tokens between chunks.

[0164] Furthermore, in the community detection module 403, the community detection algorithm is the Leiden algorithm, and the specific steps include:

[0165] Hierarchical community division is achieved through local node movement and refined partitioning;

[0166] Community summaries are generated from the bottom up. Leaf-level communities add element summaries in ascending order of node degree priority. High-level communities dynamically adjust the content according to the length of the sub-community summaries to meet token restrictions.

[0167] Furthermore, in the question-answering implementation module 404, the process of selecting the optimal answer through a multi-round voting mechanism specifically includes:

[0168] The community summary is randomly shuffled and divided into blocks, and intermediate answers and usefulness scores are generated through the Map-Reduce mechanism;

[0169] The intermediate answers are summarized in descending order of scores, and the LLM is forced to follow the reasoning process of "problem analysis → information filtering → content integration → logical arrangement" by combining the thinking chain framework;

[0170] The optimal answer is generated through multiple rounds of voting, ensuring that the answer can be traced back to knowledge graph nodes and is interpretable.

[0171] Another embodiment of this application provides a computer-readable storage medium storing computer-executable instructions for performing the above-described... Figure 1 The method for constructing a knowledge graph of key green carbon reduction technologies is shown.

[0172] This application innovatively improves upon the traditional Graph RAG method by introducing prompt tuning and chain-of-thought (CoT) mechanisms, combined with LLM (Limited Learning Model) to automate the extraction and integration of knowledge in the field of green carbon reduction. The constructed knowledge graph systematically integrates the core knowledge system, technological development trajectory, disciplinary evolution trends, and cutting-edge research dynamics of this field. Simultaneously, an intelligent question-answering system is built based on the knowledge graph of key green carbon reduction technologies, in conjunction with the CoT and voting mechanisms, effectively improving the accuracy and comprehensiveness of knowledge retrieval and the usability of the knowledge graph. This invention can provide effective decision support for reviewing the current state of technological development and analyzing future trends in the field of green carbon reduction, thus contributing to scientific research innovation and industrial development in this field.

[0173] Another aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for constructing a knowledge graph of key green carbon reduction technologies.

[0174] Specifically, the processor can be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0175] Specifically, the processor connects to the memory via a bus, which may include a path for transmitting information. The bus can be a PCI bus or an EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0176] The memory may be ROM or other types of static storage devices that can store static information and instructions, RAM or other types of dynamic storage devices that can store information and instructions, or EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0177] Optionally, the memory stores the code of the computer program that executes the solution of this application, and the execution is controlled by the processor. The processor executes the application code stored in the memory to implement the function of the intelligent question-and-answer system for a knowledge graph of key green carbon reduction technologies provided in the embodiment shown in Figure 4.

[0178] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0180] The above is a detailed description of the preferred embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for constructing a knowledge graph of key green carbon reduction technologies, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data; The entity relationship extraction task is performed by driving the large language model through domain-adapted prompt templates, and the entity recognition and relationship extraction are optimized by combining the thinking chain technology. Knowledge graph elements and element summaries are generated based on the standardized resource base. The knowledge graph and element summary are hierarchically divided into communities based on the community detection algorithm. A structured summary is generated for each community module, and a multi-granularity information mapping is constructed between the source document data layer, the knowledge graph layer and the community summary layer. Based on the thought chain-based prompting engineering framework and hierarchical graph index architecture, and using the structured summary of the community module, a comprehensive answer integrating domain knowledge is generated using a large language model, and the optimal global answer is selected through a multi-round voting mechanism.

2. The method according to claim 1, characterized in that, The data cleaning step in the step of acquiring multi-source heterogeneous data and cleaning and structuring the multi-source heterogeneous data to establish a standardized resource library includes: Text extraction and noise removal are performed on HTML and PDF documents in the multi-source heterogeneous data. Long documents are segmented using a text chunking strategy, with each text chunk having a capacity of 600 tokens and an overlap area of ​​100 tokens between chunks.

3. The method according to claim 1, characterized in that, The step of generating knowledge graph elements and element summaries by driving a large language model to perform entity relation extraction tasks through domain-adapted prompt templates, and combining mind chain technology to optimize entity recognition and relation extraction, specifically includes the following steps based on the standardized resource library: Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators; The prompt templates containing multi-step reasoning are designed using the thinking chain technology, combined with a small number of examples to guide a large language model, and entity relationships are extracted based on a well-established standardized resource library; The extracted entity, relation, and covariate information are summarized by elements to construct a knowledge graph in the form of an isomorphic undirected weighted graph.

4. The method according to claim 1, characterized in that, The step of performing hierarchical community division of the knowledge graph and element summary based on the community detection algorithm, generating a structured summary for each community module, and constructing a multi-granularity information mapping between the source document data layer, the knowledge graph layer, and the community summary layer, specifically includes the following steps: In the Leaf Layer community, element digests are added in ascending order of node degree priority until the token limit is reached; In a high-level community, if the total length of all element summaries within the community does not exceed the token limit, then the element summaries within the high-level community are added in the original order. If the total length of all element summaries within the community exceeds the token limit, then the sub-community element summaries are sorted from longest to shortest, and the shorter sub-community summaries replace the corresponding longer element summaries, gradually reducing the number of tokens until the token limit is met.

5. The method according to claim 1, characterized in that, The steps of using the thought chain-based prompting engineering framework and hierarchical graph index architecture, generating comprehensive answers that integrate domain knowledge based on the structured summary of the community module, and selecting the optimal answer through a multi-round voting mechanism, specifically include: The structured summary of the community is randomly shuffled and divided into several community summary blocks; Intermediate answers are summarized in descending order of scores. The LLM is forced to follow the reasoning process of "problem analysis → information filtering → content integration → logical arrangement" by combining the thinking chain framework and generating intermediate answers and usefulness scores based on the summary blocks of each community. Filter out intermediate answers with a score of 0, and generate a global answer by sorting the intermediate answers from each community in descending order of their usefulness scores, until the token limit is reached.

6. A knowledge graph-based intelligent question-and-answer system for key green carbon reduction technologies, characterized in that, include: The data processing module is used to acquire multi-source heterogeneous data and clean and structure the multi-source heterogeneous data, and establish a standardized resource library based on the processed multi-source heterogeneous data. The knowledge graph construction module is used to drive the large language model to perform entity relation extraction tasks through domain-adapted prompt templates, optimize entity recognition and relation extraction by combining thinking chain technology, and generate knowledge graph elements and element summaries based on the standardized resource library. The community detection module can perform hierarchical community division of the knowledge graph and element summary based on the community detection algorithm, generate a structured summary for each community module, and construct a multi-granularity information mapping between the source document data layer, the knowledge graph layer and the community summary layer. The question-answering implementation module is able to generate comprehensive answers that integrate domain knowledge based on the prompting engineering framework of the thinking chain and the hierarchical graph index architecture, and based on the structured summary of the community module, using a large language model, and select the optimal global answer through a multi-round voting mechanism.

7. The system according to claim 6, characterized in that, In the data processing module, data cleaning includes: Text extraction and noise removal are performed on HTML and PDF documents in the multi-source heterogeneous data. Long documents are segmented using a text chunking strategy, with each text chunk having a capacity of 600 tokens and an overlap area of ​​100 tokens between chunks.

8. The system according to claim 6, characterized in that, In the knowledge graph construction module, the process of generating knowledge graph elements and element summaries specifically includes: Define a list of specific entity types in the field of green carbon reduction, including technology, policy, material, technology system, energy source, and performance indicators; The prompt templates containing multi-step reasoning are designed using the thinking chain technology, combined with a small number of examples to guide a large language model, and entity relationships are extracted based on a well-established standardized resource library; The extracted entity, relation, and covariate information are summarized by elements to construct a knowledge graph in the form of an isomorphic undirected weighted graph.

9. The system according to claim 6, characterized in that, In the community detection module, the process of generating a structured summary for each community module and constructing a multi-granularity information mapping between the source document data layer, knowledge graph layer, and community summary layer specifically includes: In the Leaf Layer community, element digests are added in ascending order of node degree priority until the token limit is reached; In a high-level community, if the total length of all element summaries within the community does not exceed the token limit, then the element summaries within the high-level community are added in the original order. If the total length of all element summaries within the community exceeds the token limit, then the sub-community element summaries are sorted from longest to shortest, and the shorter sub-community summaries replace the corresponding longer element summaries, gradually reducing the number of tokens until the token limit is met.

10. The system according to claim 6, characterized in that, The question-and-answer implementation module includes a multi-round voting mechanism to select the optimal answer, specifically comprising: The structured summary of the community is randomly shuffled and divided into several community summary blocks; Intermediate answers are summarized in descending order of scores. The LLM is forced to follow the reasoning process of "problem analysis → information filtering → content integration → logical arrangement" by combining the thinking chain framework and generating intermediate answers and usefulness scores based on the summary blocks of each community. Filter out intermediate answers with a score of 0, and generate a global answer by sorting the intermediate answers from each community in descending order of their usefulness scores, until the token limit is reached.

11. A computer-readable storage medium, characterized in that, The device stores computer program instructions that, when executed by a processor, implement the method described in any one of claims 1-5.

12. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory, wherein the processor executes the program to implement the steps of any one of claims 1-5.

Images