Intelligent question and answer method and device based on multi-source ESG knowledge base and agent reasoning

Abstract

This application discloses an intelligent question-answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning, comprising: constructing a multi-source environmental and social governance knowledge base; determining the complexity of environmental and social governance issues through a first agent node; if the complexity is low, generating and outputting a fast answer; otherwise, breaking down the issues through a second agent node to determine the knowledge type and generate a recall strategy; concurrently retrieving relevant knowledge through a third agent node according to the recall strategy to obtain initial retrieval results; and determining the sufficiency of evidence based on the initial retrieval results through a fourth agent node; if the evidence is insufficient, triggering supplementary retrieval; otherwise, fusing the retrieved multi-source evidence through a fifth agent node to generate a structured answer and outputting the structured answer and its cited sources. This application can achieve accurate and efficient structured intelligent question answering in the ESG field, improving the completeness and reliability of the question-answering results.

Description

Technical Field

[0001] This application relates to the field of computer artificial intelligence technology, specifically to an intelligent question answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning. Background Technology

[0002] With the continued advancement of global sustainable development strategies, corporate environmental, social, and governance (ESG) information disclosure, compliance supervision, and professional analysis have become core tasks in capital markets, corporate management, and investment research. The diverse and heterogeneous ESG knowledge sources, including corporate ESG reports, regulatory guidelines, industry indicator systems, and market rating data, are constantly enriching the market. The demand for ESG-related Q&A has also expanded from basic information queries to complex professional scenarios such as comprehensive analysis and compliance judgments, placing higher demands on the professionalism, accuracy, and practicality of intelligent question-answering technology.

[0003] Current ESG intelligent question answering technologies generally employ single-document retrieval and traditional retrieval enhancement schemes, which can only achieve matching retrieval from a single data source. They struggle to effectively integrate and collaboratively utilize multi-source, heterogeneous ESG knowledge. Furthermore, they cannot adequately decompose and accurately match requirements when faced with highly complex ESG questions, ultimately resulting in incomplete and inaccurate question-and-answer results that are unsuitable for practical application scenarios of ESG professional question answering. In summary, existing ESG intelligent question answering technologies suffer from insufficient multi-source knowledge collaboration capabilities and low efficiency in intelligent reasoning for complex problems, leading to incomplete and inaccurate question-and-answer results.

[0004] The preceding description is intended to provide general background information and does not necessarily constitute prior art. Summary of the Invention

[0005] This application provides an intelligent question-answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning, which can achieve accurate and efficient structured intelligent question answering in the ESG field and improve the completeness and reliability of the question-answering results.

[0006] In a first aspect, embodiments of this application provide an intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning, including: Construct a multi-source environmental and social governance knowledge base, which includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. The system receives environmental and social governance questions input by users, and determines the complexity of the environmental and social governance questions through the first intelligent agent node. If the complexity is low, a quick answer is generated and output; if the complexity is high, the problem information to be broken down is output. The second intelligent agent node receives the problem information to be decomposed, decomposes the highly complex environmental and social governance problem, determines the type of knowledge required to answer the environmental and social governance problem, generates a recall strategy based on the knowledge type, and outputs the recall strategy. The third intelligent agent node receives the recall strategy and, based on the recall strategy, concurrently retrieves knowledge related to the environmental and social governance problem from the multi-source environmental and social governance knowledge base to obtain initial search results. The fourth intelligent agent node performs an evidence sufficiency judgment based on the initial search results. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, multi-source evidence to be fused is output. The fifth intelligent agent node receives the multi-source evidence to be fused, fuses the retrieved multi-source evidence, generates a structured answer, and outputs the structured answer and its cited sources.

[0007] Secondly, embodiments of this application provide an intelligent question-answering device based on a multi-source ESG knowledge base and agent reasoning, comprising: The knowledge base construction module is used to construct a multi-source environmental and social governance knowledge base, which includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. The first reasoning module is used to receive environmental and social governance questions input by the user, and to determine the complexity of the environmental and social governance questions through the first intelligent agent node. If the complexity is low, a quick answer is generated and output; if the complexity is high, the question information to be broken down is output. The second reasoning module is used to receive the problem information to be decomposed through the second intelligent agent node, decompose the highly complex environmental and social governance problem, determine the knowledge type required to answer the environmental and social governance problem, generate a recall strategy based on the knowledge type, and output the recall strategy. The third reasoning module is used to receive the recall strategy through the third intelligent agent node, and according to the recall strategy, concurrently retrieve knowledge related to the environmental and social governance problem from the multi-source environmental and social governance knowledge base to obtain initial search results; The fourth reasoning module is used to determine the sufficiency of evidence based on the initial search results through the fourth intelligent agent node. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, the multi-source evidence to be fused is output. The fifth reasoning module is used to receive the multi-source evidence to be fused through the fifth intelligent agent node, fuse the retrieved multi-source evidence, generate a structured answer, and output the structured answer and its cited sources.

[0008] This application provides an intelligent question-answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning. First, a multi-source ESG knowledge base is constructed to provide a comprehensive ESG knowledge foundation for the entire question-answering process, offering sufficient information support for subsequent knowledge retrieval and answer generation. Then, multiple agent nodes sequentially complete the entire process, including question complexity assessment, decomposition of high-complexity questions, concurrent retrieval according to a recall strategy, evidence sufficiency verification, and multi-source evidence fusion and structured output. The first agent node distinguishes the complexity of questions, directly generating rapid answers for low-complexity questions to simplify the processing and improve response time. The process involves several steps: First, a high-complexity problem is broken down to clarify the core solution direction. Then, a second intelligent agent node determines the required knowledge type based on the decomposition results and generates a matching recall strategy, defining a path for accurate retrieval. A third intelligent agent node concurrently retrieves relevant knowledge from a multi-source ESG knowledge base according to the recall strategy, efficiently acquiring information matching the problem. A fourth intelligent agent node assesses the sufficiency of evidence in the retrieval results, ensuring the answer has sufficient factual basis and avoiding bias due to insufficient information. Finally, a fifth intelligent agent node integrates multi-source evidence and generates a structured answer with cited sources, making the output logically clear and verifiable. In summary, this application, through comprehensive support from a multi-source knowledge base and collaborative reasoning by multiple intelligent agents, achieves adaptable processing of ESG problems of varying complexity, balancing answering efficiency and accuracy, ultimately forming rigorous and standardized ESG intelligent question-answering results that can stably adapt to the answering needs of various ESG problems. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is an application environment diagram of the intelligent question answering method based on multi-source ESG knowledge base and agent reasoning provided in the embodiments of this application; Figure 2This is a flowchart illustrating the intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning provided in an embodiment of this application. Figure 3 This is a flowchart illustrating the ESG intelligent question-answering method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the architecture of the ESG intelligent question-answering system provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the intelligent question-answering device based on a multi-source ESG knowledge base and agent reasoning provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0011] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with those detailed in the appended claims or with some aspects of this application.

[0012] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover descriptions such as non-exclusive inclusion, so that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.

[0013] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0014] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustrative purposes and has no specific meaning in itself. Therefore, "module," "part," or "unit" may be used interchangeably.

[0015] To address the aforementioned technical problems and overcome the shortcomings of existing technologies, this application provides an intelligent question-answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning, which can achieve accurate and efficient structured intelligent question answering in the ESG field and improve the completeness and reliability of the question-answering results.

[0016] Figure 1 This is an application environment diagram of an intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning in one embodiment. (Refer to...) Figure 1 This intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning is applied to an intelligent question-answering system based on a multi-source ESG knowledge base and agent reasoning. The intelligent question-answering system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. Server 120 is configured to execute the aforementioned intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning, including: constructing a multi-source environmental and social governance knowledge base; receiving environmental and social governance questions input by users, determining the complexity of the environmental and social governance questions through a first agent node; if the complexity is low, generating and outputting a quick answer; if the complexity is high, outputting information about the questions to be broken down; receiving the information about the questions to be broken down through a second agent node, breaking down high-complexity environmental and social governance questions to determine the types of knowledge required to answer the environmental and social governance questions, and generating a recall strategy based on the knowledge types. Output a recall strategy; receive the recall strategy through a third intelligent agent node, and concurrently retrieve knowledge related to environmental and social governance issues from a multi-source environmental and social governance knowledge base according to the recall strategy to obtain initial search results; the fourth intelligent agent node performs evidence sufficiency judgment based on the initial search results. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, the multi-source evidence to be fused is output; the fifth intelligent agent node receives the multi-source evidence to be fused, fuses the retrieved multi-source evidence, generates a structured answer, and outputs the structured answer and its cited sources.

[0017] Please see Figure 2 , Figure 2 This is a flowchart illustrating an intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning, provided in an embodiment of this application. This embodiment primarily uses the application of this intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning to a computer device as an example. Specifically, the intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning provided in an embodiment of this application may include the following steps: S10. Construct a multi-source environmental and social governance knowledge base, which includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. Specifically, for step S10, the first step is to establish a diversified and comprehensive ESG knowledge storage system. This multi-source environmental and social governance knowledge base includes: a first knowledge sub-base for storing corporate ESG reporting data; a second knowledge sub-base for storing semantically enhanced ESG indicator data; and a third knowledge sub-base for storing ESG regulations and knowledge graph data. By integrating at least one of these knowledge sub-bases, a multi-dimensional ESG knowledge base is formed, providing comprehensive data source support for subsequent question answering. For example, the first knowledge sub-base can be built separately to store corporate ESG reporting data, or the first and third knowledge sub-bases can be combined, incorporating both corporate disclosure data and regulatory rules data to meet the knowledge supply needs of basic question answering.

[0018] S20. Receive the environmental and social governance problem input by the user, and determine the complexity of the environmental and social governance problem through the first intelligent agent node. If it is determined to be low complexity, generate a quick answer and output it. If it is determined to be high complexity, output the problem information to be broken down. Specifically, in step S20, the system first receives ESG-related questions input by the user. The first intelligent agent node then determines the complexity of the question: if the question is low-complexity, a quick answer is generated and output directly based on the knowledge base, completing the question-and-answer process; if the question is high-complexity, no answer is generated directly, but the question is only organized into information to be broken down and output, providing a foundation for subsequent processing of complex questions. For example, a user asking "What dimensions does ESG include?" is a low-complexity question, and the first intelligent agent node directly and quickly outputs the answer; a user asking "Does a company's ESG management meet regulatory requirements?" is a high-complexity question, and the first intelligent agent node only outputs the information to be broken down for that question.

[0019] S30. Receive the problem information to be decomposed through the second intelligent agent node, decompose the highly complex environmental and social governance problem to determine the knowledge type required to answer the environmental and social governance problem, generate a recall strategy based on the knowledge type, and output the recall strategy. Specifically, in step S30, the second agent node receives the problem information to be decomposed from the first agent node, performs decomposition analysis on the high-complexity ESG problem, clarifies the type of knowledge required to answer the problem, formulates a matching recall strategy based on the determined knowledge type, and finally outputs the generated recall strategy to define the execution direction for subsequent knowledge retrieval. This transforms the complex problem into a feasible retrieval instruction, improving the targeting of subsequent knowledge acquisition. S40. Receive the recall strategy through the third intelligent agent node, and according to the recall strategy, concurrently retrieve knowledge related to environmental and social governance issues from the multi-source environmental and social governance knowledge base to obtain initial search results; Specifically, in step S40, the third agent node receives the recall strategy generated by the second agent node and strictly follows this strategy to synchronously retrieve multiple knowledge sources from the pre-built multi-source ESG knowledge base, quickly obtaining knowledge content related to the user's ESG issues, and finally compiling the initial retrieval results. This concurrent retrieval method allows for the simultaneous retrieval of multi-dimensional knowledge, significantly improving the efficiency of knowledge acquisition.

[0020] S50. The fourth intelligent agent node performs a sufficiency judgment on the evidence based on the initial search results. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, the multi-source evidence to be fused is output. Specifically, in step S50, the fourth agent node, based on the initial search results obtained from the third agent node, evaluates whether the evidence supporting the answer to the question is sufficient. If the evidence is deemed insufficient, a supplementary search instruction is generated and sent back to the third agent node, triggering a new round of supplementary search. If the evidence is deemed sufficient, the valid evidence obtained from the multi-source search is organized into multi-source evidence to be fused and output. Through evidence verification, the completeness of the information in answering the question is ensured, and deviations in the answer are avoided.

[0021] S60. Receive the multi-source evidence to be fused through the fifth intelligent agent node, fuse the retrieved multi-source evidence, generate a structured answer, and output the structured answer and its cited sources; Specifically, in step S60, the fifth intelligent agent node receives the multi-source evidence to be fused output by the fourth intelligent agent node, integrates and processes the knowledge evidence from different sources, generates a logically clear structured answer based on the fused information, and outputs the structured answer along with the corresponding cited sources to ensure the standardization and traceability of the answer.

[0022] This embodiment uses a multi-source ESG knowledge base as its knowledge support. Through multiple intelligent agent nodes, it sequentially completes the following tasks: problem complexity judgment, high-complexity problem decomposition, multi-source concurrent retrieval, evidence sufficiency verification, multi-source evidence fusion, and structured output. It can adapt to ESG problem processing of different complexities, and balances the efficiency of answering with the accuracy and completeness of the answers, ultimately achieving standardized and reliable ESG intelligent question answering.

[0023] Furthermore, in some embodiments, the construction method of the first knowledge sub-base in step S10 may specifically include: S101A. Obtain the corporate environmental and social governance report text, parse the report text and divide it into multiple paragraphs; Specifically, for step S101A, the first step is to collect the environmental and social governance (ESG) report text disclosed by the company. The report text undergoes format parsing and content extraction. After removing redundant formatting information, the complete report text is divided into multiple independent paragraph units according to the text's content logic, chapter structure, or natural expression boundaries. Each paragraph retains relatively complete ESG disclosure information, providing a basic text unit for subsequent summary generation and vector conversion. For example, a listed company's annual ESG report includes content on environmental emission reduction, employee rights, and supply chain management. This complete report is divided into 40 independent paragraphs, each corresponding to a specific ESG disclosure item.

[0024] S102A. Generate summary information for each paragraph based on a preset language model, convert the summary information into a first semantic vector, and construct a summary vector sub-library; Specifically, for step S102A, based on a preset language model, the core information of each segment is extracted to generate a summary of the corresponding segment. The summary retains only the key content of the segment. Then, each summary is converted into a first semantic vector, and the first semantic vectors corresponding to all segments are aggregated and stored to form a summary vector sub-library. For example, for the "Water Resource Management" segment in a company's ESG report, the language model extracts the summary "In 2025, the company's water resource consumption decreased by 10% year-on-year, and the water cycle transformation project was completed." This summary is then converted into a first semantic vector and stored in the summary vector sub-library to quickly locate the core information of the segment.

[0025] S103A. Convert each original paragraph of the report text into a second semantic vector to construct a content vector sub-library; wherein, the first semantic vector and the second semantic vector have different dimensions, the summary vector sub-library is used to quickly locate core information, and the content vector sub-library is used to provide factual details and evidence; Specifically, in step S103A, the obtained original report paragraphs are directly converted into second semantic vectors. All second semantic vectors corresponding to the original paragraphs are then aggregated and stored to form a content vector sub-library. The first and second semantic vectors are set to different dimensions. The summary vector sub-library uses concise summary information to quickly locate core information, while the content vector sub-library provides detailed factual evidence for question answering based on the complete original paragraph text. For example, the complete original paragraph of "water resource management" (including all details such as consumption data, renovation plans, implementing departments, and acceptance results) is retained. This complete original paragraph is converted into a second semantic vector with a different dimension than the first semantic vector and stored in the content vector sub-library to provide accurate factual evidence.

[0026] This embodiment establishes a two-layer vector sub-library by separating the summary from the original text and storing it in a two-dimensional vector database. This allows for the rapid location of core information in ESG reports while providing complete factual details, thus balancing the efficiency of knowledge retrieval with the accuracy of information support.

[0027] Furthermore, in some embodiments, the construction method of the second knowledge sub-base in step S10 may specifically include: S101B. Collect the names of environmental and social governance indicators and their corresponding rating criteria; Specifically, step S101B primarily provides the basic raw data for semantic enhancement processing of indicators. This involves actively collecting and organizing the names of ESG indicators under various standards and evaluation systems in the ESG field. Simultaneously, it collects and matches the core rating points corresponding to each indicator in market rating and compliance evaluation processes, mapping indicator names to rating points one-to-one to form a complete set of raw ESG indicator data. This provides standardized material for subsequent semantic expansion and vector conversion. For example, collecting the ESG indicator name "water resource consumption intensity" and simultaneously collecting its corresponding rating points, including "the requirement to disclose total water resource consumption annually, water consumption per unit of revenue, and the implementation status of water-saving measures," etc.

[0028] S102B. Generate natural language explanatory text for each environmental and social governance indicator using a language model. The natural language explanatory text includes at least one of the indicator's connotation, calculation method, or disclosure requirements. Specifically, for step S102B, the compiled indicator names and rating points are input into a preset language model. The language model then semantically expands each ESG indicator to generate a natural language explanation text. This explanation text must include at least one or more of the following: indicator connotation, calculation method, and disclosure requirements. This transforms the originally concise and semantically weak indicator information into complete, logically clear, and easily understood text content, enhancing the semantic expressive power of the indicator. For example, for the indicator "water resource consumption intensity," the language model generates the following natural language explanation text: "This indicator is used to measure the efficiency of water resource utilization in enterprise production and operation. The calculation method is the total annual water resource consumption divided by the enterprise's annual operating revenue. When disclosing, the statistical scope, investment in water-saving renovations, and the completion status of efficiency improvement targets must be clearly stated."

[0029] S103B. Combine the names of environmental and social governance indicators, rating points, and natural language explanation text, and convert them into semantic vectors to construct a semantically enhanced vector sub-library for semantic alignment between different environmental and social governance knowledge sources. Specifically, in step S103B, the name, rating criteria, and natural language explanation text corresponding to each ESG indicator are integrated and concatenated to form a complete semantically enhanced text for the indicator. This combined text is then uniformly converted into a semantic vector, and the semantic vectors corresponding to all indicators are centrally stored, ultimately constructing a semantically enhanced vector sub-library. This sub-library can be used to achieve precise semantic alignment between different ESG knowledge sources. For example, the indicator name, rating criteria, and natural language explanation text for "water resource consumption intensity" are combined into a complete text, converted into a semantic vector, and stored in the sub-library. Subsequently, semantically unified matching of this indicator can be achieved across different knowledge sources such as corporate reports, regulations, and rating data.

[0030] This embodiment significantly enriches the semantic information of ESG indicators by performing semantic enhancement processing on ESG indicators and building a dedicated vector sub-library, enabling efficient semantic alignment between different ESG knowledge sources.

[0031] Furthermore, in some embodiments, the construction method of the third knowledge sub-base in step S10 may specifically include: S101C. Collect environmental and social governance (ESG) regulations and / or ESG knowledge graph data, which includes indicators, enterprises, industries, and constraint relationships. Specifically, for step S101C, which involves acquiring the basic data source for the third knowledge sub-base, one or both types of data can be collected simultaneously, depending on actual needs. One type is various environmental and social governance (ESG) regulations, industry standards, and policy requirements applicable to enterprises; the other type is the completed ESG knowledge graph data, which contains four core related information categories: indicators, enterprises, industries, and constraint relationships. By collecting the above data, complete raw materials are provided for subsequent segmentation and vector transformation. For example, regulatory provisions related to enterprise ESG information disclosure can be collected, while ESG knowledge graph data containing the relationship between "carbon emission reduction indicators—new energy enterprises—new energy industry—emission reduction compliance constraints" can also be collected.

[0032] S102C. Following a block-segmentation strategy that prioritizes logical integrity, legal clauses or knowledge graph data are divided into blocks according to logical units so that each block maintains complete normative semantics. Specifically, for step S102C, following the principle of prioritizing logical integrity in the segmentation, meaningless fragmentation of regulatory clauses and knowledge graph data is avoided. Instead, the collected regulatory clauses or knowledge graph data are segmented based on logical units with independent, complete, and normative semantics. This approach ensures that each segment can independently carry complete regulatory requirements or knowledge graph relational logic, preventing damage to the semantic integrity and logical normativity of the content due to excessive segmentation. For example, the "complete requirements for employee rights protection disclosure" in ESG regulations is segmented as a single logical unit, without being broken down into fragmented statements; the "ESG compliance constraint relationship group of a certain manufacturing industry" in the knowledge graph is segmented as a single logical unit, preserving the complete relational logic of indicators, enterprises, industries, and constraints.

[0033] S103C. Convert each block into a semantic vector and record the structural information of each block to build a regulatory and knowledge graph vector sub-library for applying structural constraints during the recall phase. Specifically, for step S103C, each logical block is first converted into a semantic vector, and the structural information corresponding to each block is recorded, including the regulation number, clause / section, and knowledge graph node relationship identifier. The semantic vectors and accompanying structural information of all blocks are integrated and stored to form a regulation and knowledge graph vector sub-library. This sub-library can apply structural constraints based on the recorded structural information in the subsequent knowledge retrieval stage to ensure the standardization of the retrieval. For example, the "Employee Rights Protection Disclosure Requirements" block is converted into a semantic vector, and its corresponding regulation number and clause number are recorded and stored in the sub-library. During subsequent retrievals, the structural information can be used to limit the search scope and improve matching accuracy.

[0034] This embodiment processes regulatory and knowledge graph data with a priority of logical integrity. It combines vector transformation and structural information recording to build a dedicated sub-library, which can not only preserve the complete normative semantics of ESG regulations and knowledge graphs, but also impose structural constraints during retrieval, thereby improving the standardization and accuracy of knowledge retrieval.

[0035] Furthermore, in some embodiments, step S20, "receiving an environmental and social governance question input by a user, determining the complexity of the environmental and social governance question through a first intelligent agent node, generating a quick answer and outputting it if the complexity is low, and outputting the question information to be broken down if the complexity is high," may specifically include: S201. Receive user input regarding environmental and social governance issues; Specifically, in step S201, the system receives user queries related to Environmental and Social Governance (ESG) in natural language via the interactive port. This process collects and transmits the queries, using the original user queries as the basis for subsequent complexity assessments, thus providing the foundational input for the intelligent agent node's decision-making operations. For example, the system receives queries such as "What does ESG mean?" or "What are the core contents of Company A's 2024 ESG report?"

[0036] S202. The complexity of environmental and social governance issues is judged by a preset language model. If an environmental and social governance issue can be answered with only a single fact or concept, it is judged as low complexity; if an environmental and social governance issue requires the integration of multi-source knowledge, compliance judgment, or comprehensive analysis, it is judged as high complexity. Specifically, in step S202, a pre-defined language model is invoked to perform semantic analysis and difficulty assessment on the received ESG questions. The complexity of the questions is categorized according to a unified standard. If the question can be answered using only a single factual data point or a single conceptual definition, it is classified as a low-complexity question. If the question requires the integration of multi-source knowledge, compliance judgment, or comprehensive analytical reasoning to be answered, it is classified as a high-complexity question. For example, a user asking "What are Scope3 emissions?" can be answered with only a conceptual explanation and is classified as low-complexity; a user asking "Does Company A's ESG management measures comply with industry regulatory requirements?" requires integrating company data, regulatory rules, and making a compliance judgment and is classified as high-complexity.

[0037] S203. If the complexity is determined to be low, a quick answer will be generated and output from the multi-source environmental and social governance knowledge base, and the process will end. Specifically, for step S203, for low-complexity ESG questions, the system directly retrieves corresponding information from the pre-built multi-source ESG knowledge base, quickly generates a concise answer, and outputs it to the user. This concludes the question-and-answer process, eliminating the need for subsequent complex processing steps. For example, for the low-complexity question "What are the three aspects of ESG?", the system retrieves the corresponding concepts from the knowledge base and directly outputs the quick answer "environment, society, and governance," ending the question-and-answer process.

[0038] S204. If the problem is determined to be of high complexity, output the problem information to be broken down. Specifically, for step S204, for highly complex ESG questions, instead of directly generating an answer, the system organizes the original content and core query requirements of the question into question information to be broken down and outputs it, providing input for subsequent in-depth question processing. For example, for the highly complex question "Compare the ESG compliance performance of companies A and B over the past three years," the system organizes the core information of the question, outputs the question information to be broken down, and submits it to subsequent processes.

[0039] This embodiment uses a language model to determine the complexity of ESG questions and perform traffic splitting. Low-complexity questions are responded to quickly, while high-complexity questions are transferred to a subsequent decomposition process. This achieves a reasonable layering of the question-and-answer process and effectively improves the overall execution efficiency of ESG question-and-answer.

[0040] Furthermore, in some embodiments, step S30, "receiving the problem information to be decomposed through the second intelligent agent node, decomposing the highly complex environmental and social governance problem to determine the knowledge type required to answer the environmental and social governance problem, generating a recall strategy based on the knowledge type, and outputting the recall strategy," may specifically include: S301. Identify the core needs, required knowledge types, and key elements involved in environmental and social governance issues. Key elements include enterprises, industries, or indicators. Specifically, for step S301, the second intelligent agent node first performs a full-dimensional analysis of the received highly complex Environmental and Social Governance (ESG) problem through semantic decomposition, keyword extraction, and entity recognition. It first analyzes the problem statement sentence by sentence, removes modifying statements, and locates the user's core solution requirements. Then, based on the core requirements, it reverse-engineers the knowledge categories that must be relied upon to solve the problem. At the same time, through entity matching algorithms, it accurately extracts key limiting elements such as enterprises, industries, and ESG indicators involved in the problem. Finally, it fully clarifies the core requirements, required knowledge types, and key elements of the problem, providing accurate basic information for subsequent processing.

[0041] For example, regarding the question of whether a new energy company's carbon emission reduction targets in 2024 comply with domestic ESG regulatory requirements, the core requirement is first determined through semantic decomposition to judge the compliance of the company's carbon emission reduction behavior; then the required knowledge types are derived as the company's carbon emission reduction disclosure data and ESG regulatory standards; finally, key elements are extracted through entity recognition: a new energy company, the new energy industry, and carbon emission reduction targets.

[0042] S302. Determine the question type based on core needs. Question types include concept definition, business fact, compliance judgment, market opinion, or comprehensive comparative analysis. Specifically, in step S302, the second intelligent agent node will match and compare the extracted core requirements with the preset five types of ESG problem standard features one by one. According to the principle of the highest feature fit, the specific type of the current problem is determined. The five types of problem are concept definition, enterprise fact, compliance judgment, market opinion and comprehensive comparative analysis. Through standardized classification, the subsequent strategy generation has a clear classification basis.

[0043] For example, if the core need is to explain the meaning of professional terms / indicators, it is determined to be a concept definition category after matching features; if the core need is to query specific disclosed data of an enterprise, it is determined to be a corporate fact category after matching features; if the core need is to determine whether a behavior complies with rules / standards, it is determined to be a compliance judgment category after matching features; if the core need is to understand the market's evaluation of an enterprise / industry, it is determined to be a market opinion category after matching features; if the core need is to compare the ESG performance of multiple entities, it is determined to be a comprehensive comparative analysis category after matching features.

[0044] S303. Generate corresponding recall strategies based on question types, wherein the recall strategies corresponding to different question types prioritize recalling knowledge from different knowledge sub-bases; Specifically, for step S303, the second intelligent agent node generates a dedicated recall strategy for the identified types of questions according to the preset rules that bind the question type to the knowledge sub-base. The core is to clarify which type of knowledge sub-base to retrieve knowledge from when different question types are retrieved, solidify the retrieval direction and knowledge source priority into executable retrieval rules, and finally form a standardized recall strategy and output it.

[0045] For example, if the question is identified as a concept definition question, a recall strategy is generated that prioritizes retrieving sub-bases of knowledge related to indicators and regulations; if the question is identified as a business fact question, a recall strategy is generated that prioritizes retrieving sub-bases of knowledge related to corporate ESG reports; if the question is identified as a compliance judgment question, a recall strategy is generated that prioritizes retrieving sub-bases of knowledge related to regulations; if the question is identified as a market opinion question, a recall strategy is generated that prioritizes retrieving sub-bases of knowledge related to market ratings; and if the question is identified as a comprehensive comparative analysis question, a recall strategy is generated that integrates multiple sub-bases of knowledge related to enterprises, indicators, and markets.

[0046] This embodiment analyzes and categorizes highly complex ESG questions from all dimensions, generates exclusive recall strategies that match the question type, accurately identifies the source and direction of knowledge retrieval, makes subsequent knowledge acquisition more targeted, and improves the rationality and accuracy of ESG question answering.

[0047] Furthermore, in some embodiments, step S303, "generating a corresponding recall strategy based on the problem type," may specifically include: S3031. If the problem type is a concept definition type, data should be retrieved from the second or third knowledge sub-base first. S3032. If the question type is enterprise fact type, data should be retrieved from the first knowledge sub-base first; S3033. If the problem type is compliance judgment, data should be retrieved from the third knowledge sub-base first, and semantic alignment should be performed in combination with the second knowledge sub-base, and then the enterprise disclosure should be verified through the first knowledge sub-base. S3034. If the question type is market opinion, data should be retrieved first from the market rating and online supplementary sub-library, supplemented by data from the first knowledge sub-library; S3035. If the problem type is a comprehensive comparative analysis, then the second knowledge sub-base, the first knowledge sub-bases of multiple enterprises, and market rating data are integrated for recall.

[0048] Specifically, when the problem is determined to be a concept definition, the priority target for knowledge retrieval is locked to the second knowledge sub-base (ESG indicator semantic enhancement data) or the third knowledge sub-base (ESG regulations and knowledge graph data) according to the preset retrieval rules. Knowledge content related to indicator definitions and criterion interpretations is retrieved directly from the above sub-bases to quickly obtain authoritative and standard concept interpretation information.

[0049] When the issue is determined to be a factual issue for the enterprise, the priority target for knowledge retrieval is locked to the first knowledge sub-base (enterprise ESG report data). The original text of the corresponding enterprise's ESG disclosure, specific data, implementation measures and other factual information are directly retrieved from this sub-base to accurately match the enterprise's specific ESG situation.

[0050] When the issue is determined to be a compliance judgment, a three-tiered progressive recall logic is executed: First, compliance judgment basis such as legal provisions and regulatory requirements are retrieved from the third knowledge sub-base; then, the semantic data of indicators from the second knowledge sub-base is combined to complete the semantic unification matching between legal requirements and enterprise indicators; finally, the enterprise disclosure data from the first knowledge sub-base is used to verify whether the enterprise's actual behavior complies with compliance requirements, forming a complete compliance judgment knowledge chain.

[0051] When the issue is determined to be a market opinion, the priority targets for knowledge retrieval are market ratings and supplementary online sub-libraries, retrieving market institution rating conclusions, industry opinions, and other content; at the same time, corporate ESG report data from the first knowledge sub-library are used as auxiliary support, combined with market evaluations and actual corporate disclosures, to reconstruct the complete basis for market opinions.

[0052] When the problem is determined to be a comprehensive comparative analysis, the multi-database integration recall logic is executed: the indicator standards of the second knowledge sub-base, the disclosure data of the first knowledge sub-base corresponding to multiple companies, and the market rating data are integrated and retrieved to obtain the unified standards, factual data of each subject, and market evaluation data required for comparison, so as to support multi-dimensional comprehensive comparative analysis.

[0053] This embodiment matches differentiated multi-database recall rules for different types of ESG issues, making knowledge recall more aligned with problem-solving needs and improving the accuracy and adaptability of knowledge acquisition.

[0054] Furthermore, in some embodiments, step S40, "receiving the recall strategy through a third intelligent agent node and, according to the recall strategy, concurrently retrieving knowledge related to environmental and social governance issues from a multi-source environmental and social governance knowledge base to obtain initial search results," may specifically include: S401. Based on the recall strategy, call the retrieval interfaces of each sub-base in the multi-source environmental and social governance knowledge base in parallel; Specifically, for step S401, the third intelligent agent node first parses the received recall strategy, clarifies the type of knowledge sub-base to be called, the search scope and execution requirements, and then sends search instructions to the corresponding target sub-bases in the multi-source ESG knowledge base in a multi-threaded parallel scheduling manner. This directly triggers the synchronous operation of the search interfaces of each sub-base, without waiting for the search of a single sub-base to be completed before starting the next one. By replacing serial search with parallel processing, the overall knowledge acquisition time is greatly shortened.

[0055] S402. Assess the relevance of the search results for each sub-database and perform preliminary filtering based on the scores; Specifically, for step S402, for the original search results returned independently by each sub-database, a semantic feature comparison method is used for scoring. The user's original ESG question is semantically matched with the text content of each search result, calculating the degree of association between their core expressions, key information, and query intent. A quantitative relevance score of 0-100 is generated, with higher scores indicating a higher degree of matching between the search result and the user's question. A unified scoring threshold is pre-set; search results with relevance scores below the threshold are judged as low-relevance, unreliable, and invalid information and are directly removed. Only search results with scores reaching or exceeding the threshold and highly matching the question are retained, reducing the amount of data processed subsequently.

[0056] S403. Perform deduplication on the filtered search results to obtain the initial search results; Specifically, for step S403, firstly, for all the search results that have undergone preliminary filtering, extract the unique feature information such as the core text, key data, and core expressions of each result; then, perform feature cross-comparison on the results returned by different sub-databases to identify redundant data with completely identical content, repeated core information, and highly similar expressions; subsequently, retain one valid data and remove the remaining duplicate information to avoid duplicate content interfering with subsequent processes. Finally, integrate and summarize all the deduplicated valid and highly relevant search results to form a well-organized initial search result.

[0057] This embodiment improves retrieval efficiency by calling multiple sub-library retrieval interfaces in parallel, uses semantic similarity scoring to filter highly relevant information and text feature comparison to remove duplicate data, and finally obtains efficient, accurate and non-redundant initial retrieval results, providing reliable data support for subsequent question-answering processing.

[0058] Furthermore, in some embodiments, step S402, "scoring the relevance of the search results for each sub-database and performing preliminary filtering based on the scores," may specifically include: The system rewrites user-input environmental and social governance questions, generates multiple rewritten queries with different semantic perspectives, and uses these rewritten queries for retrieval. And / or, introduce a re-ranking model into the preliminary search results to re-rank them based on semantic relevance, source authority, or content completeness.

[0059] Specifically, based on the user's original input regarding environmental and social governance issues, the original question is rewritten through methods such as synonym conversion, core element expansion, semantic perspective switching, and expression optimization. This generates multiple rewritten queries with the same core demand but different expressions and semantic perspectives. These rewritten queries are then used as new search conditions, and search operations are performed on them separately. This addresses the problem of the original question being too simplistic and lacking sufficient search coverage, recalling relevant knowledge from more dimensions and avoiding the omission of key information. For example, if the user's original question is "A company's carbon emission reduction measures in 2024," after query rewriting, multiple query statements from different perspectives are generated, including "A company's specific greenhouse gas emission reduction plan for 2024," "A company's low-carbon and environmental protection implementation measures in 2024," and "A company's ESG carbon emission control measures in 2024." These rewritten queries are then used to conduct searches, recalling relevant information from more dimensions.

[0060] All results obtained from the initial search are input into a pre-defined re-ranking model. The model uses semantic relevance, source authority, and content completeness as core evaluation dimensions to quantitatively assess each search result: semantic relevance measures the closeness of the match between the result and the user's question; source authority determines the professionalism and credibility of the knowledge source; and content completeness determines whether the result contains complete answer information. Based on the comprehensive evaluation score, the model re-ranks the search results in descending order, placing results that are more relevant to the question, more authoritative, and more complete at the top, thus optimizing the ranking structure of the search results. For example, if the initial search results include official corporate ESG reports, industry analysis articles, and online self-media information, the re-ranking model, after evaluation, will place the corporate ESG report with high semantic matching and authoritative source at the top, followed by the industry analysis article, and finally the online self-media information, thus completing the re-ranking of the search results.

[0061] This embodiment broadens the retrieval coverage by query rewriting and optimizes the ranking of retrieval results by combining a re-ranking model, effectively improving the comprehensiveness and reliability of retrieval results.

[0062] Furthermore, in some embodiments, step S50, "the fourth intelligent agent node performs an evidence sufficiency judgment based on the initial search results; if the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node; if the evidence is deemed sufficient, the multi-source evidence to be fused is output," may specifically include: S501. Obtain the user's original target, the decomposition results output by the second agent node, the initial search results, and the search operation information that has been executed; Specifically, for step S501, the fourth agent node first comprehensively collects all key process information generated in this question-and-answer process. This includes the original core objective of the user's question, the problem analysis results after the second agent node decomposes the highly complex problem, the initial search results obtained after the third agent node completes concurrent retrieval, and the specific information of all retrieval operations previously executed by the system. These four types of information are then fully integrated as the basic input data for determining the sufficiency of evidence, ensuring that the judgment process relies on the entire process information without any omissions of key steps.

[0063] For example, the user's original goal is to check whether a company's wastewater discharge in 2024 is compliant; the problem breakdown results in two types of knowledge: the company's wastewater discharge data and environmental regulations; the initial search results only contain the company's discharge data; the search operation has only retrieved the first knowledge sub-base, and all of the above information has been completely obtained by the fourth intelligent agent node.

[0064] S502. Based on the preset sufficiency evaluation dimensions, the initial search results are judged for evidence sufficiency. The sufficiency evaluation dimensions include the coverage of key knowledge types, the completeness of evidence, and the degree of fit between the search results and the user's goals. Specifically, in step S502, the fourth agent node comprehensively evaluates the initial search results according to three pre-defined core evaluation dimensions. This includes determining whether the search results cover all knowledge types required to answer the question, with no missing core knowledge categories; whether the information in the existing search results is complete, with no missing key data, core clauses, or important expressions; and whether the search results align with the core requirements of the user's original question, without any invalid information deviating from the query direction. Through cross-evaluation across these three dimensions, the node ultimately determines whether the existing evidence is sufficient to support the generation of an accurate and complete answer.

[0065] For example, compliance verification requires two types of knowledge: enterprise data and regulatory data. The initial search only covers enterprise data, which is insufficient in terms of coverage of key knowledge types. Without regulatory provisions, a compliance determination cannot be made, and the completeness of evidence is lacking. The search results only provide enterprise data, which does not align with the core objectives of compliance verification, and the overall judgment is that the evidence is insufficient.

[0066] S503. If the evidence is deemed insufficient, a supplementary search instruction is generated and returned to the third agent node to trigger a supplementary search. Specifically, in step S503, after determining that the initial search evidence is insufficient, the fourth agent node generates a targeted supplementary search instruction based on issues such as the missing knowledge type and the uncovered search direction, clarifying the knowledge source, search scope, core keywords, and other content that need to be supplemented; then the instruction is sent back to the third agent node, directly triggering a new round of supplementary search operations to fill the evidence gap.

[0067] For example, if evidence related to environmental regulations is missing, a supplementary search instruction is generated to "search for the corresponding wastewater discharge regulations in the third knowledge sub-base" and returned to the third intelligent agent node to trigger the supplementary search.

[0068] S504. If the evidence is deemed sufficient, output the multi-source evidence to be fused; Specifically, for step S504, when it is determined that the initial search results meet the standards in all three evaluation dimensions and the existing evidence is sufficient to completely and accurately answer the user's question, the fourth intelligent agent node organizes and sorts out the initial search results, removes invalid and redundant information, retains all valid multi-source evidence, organizes it into standardized multi-source evidence to be fused and outputs it, providing reliable input for subsequent evidence integration and answer generation.

[0069] For example, the initial search results fully cover the enterprise's wastewater discharge data and corresponding environmental regulations, the information is complete and fits the compliance verification objectives, the evidence is sufficient, and after sorting, multi-source evidence to be integrated is output.

[0070] This embodiment accurately identifies information gaps by judging the sufficiency of evidence from multiple dimensions. If the evidence is insufficient, it triggers a supplementary search; if the evidence is sufficient, it outputs well-organized evidence, providing reliable information support for the subsequent generation of answers.

[0071] Furthermore, in some embodiments, step S502, "based on preset sufficiency evaluation dimensions, performs an evidence sufficiency judgment on the initial search results," further includes: Semi-structured constraints are introduced, and a language model is used to perform a self-assessment of a pre-set checklist. The checklist includes whether the company's disclosed information has been retrieved, whether authoritative regulations have been referenced, and whether at least one piece of factual evidence exists.

[0072] Specifically, step S502 also includes incorporating semi-structured constraints into the evidence sufficiency assessment process to enhance the standardization and objectivity of the assessment through standardized verification methods. The specific steps are as follows: Based on the existing evidence sufficiency assessment, a semi-structured judgment constraint that is not rigidly coded and is flexible and executable is added. Instead of adopting a fixed and rigid procedural judgment logic, a standardized judgment basis is built in the form of a checklist, so that evidence verification is more in line with the actual needs of ESG professional Q&A.

[0073] A core checklist is pre-defined, which includes three basic and key check items: whether the company's disclosed information has been retrieved, whether authoritative regulations have been referenced, and whether there is at least one piece of factual evidence. These three items are the core check points to ensure that the ESG Q&A results have basic credibility.

[0074] Input the initial search results, completed search operations, and other relevant information into the preset language model. The model will then independently check and judge each of the three items on the checklist, and give a clear conclusion of "satisfied" or "not satisfied" for each item, forming a complete self-evaluation result of the checklist.

[0075] The self-assessment results of the language model checklist are used as an important reference for judging the sufficiency of evidence. Combined with the original evaluation dimensions, this further improves the rigor of the judgment results and avoids judgment bias due to the lack of basic evidence.

[0076] For example, regarding the question of "whether a company's 2024 emissions were compliant," the language model performs a self-assessment based on the checklist: ① Has the company's disclosed information been retrieved?: Satisfied (relevant data from the company's ESG report has been retrieved); ② Has authoritative regulations been consulted?: Not satisfied (no corresponding environmental regulations were retrieved); ③ Is there at least one piece of factual evidence?: Satisfied (specific data on the company's emissions is available). Based on this self-assessment, it can be clearly determined that the current evidence lacks a fundamental basis and is insufficient.

[0077] This embodiment introduces semi-structured constraints and relies on a language model to complete a self-assessment of a preset list, making the judgment of evidence sufficiency more standardized and objective, and effectively ensuring the integrity and reliability of the basic evidence for ESG question answering.

[0078] Furthermore, in some embodiments, the method further includes: when a supplementary search is triggered, setting a maximum number of iterations; when the number of iterations for the supplementary search exceeds the maximum number of iterations, the fourth agent node forcibly outputs the current multi-source evidence as the multi-source evidence to be fused.

[0079] Specifically, when the fourth agent node determines that the evidence is insufficient and formally triggers a supplementary search operation, the system pre-sets a maximum number of iterations for the supplementary search. This maximum number is the upper limit for the number of times the supplementary search can be executed repeatedly, thus limiting the number of times the supplementary search can be repeated and preventing the process from stalling due to endless supplementary searches. The system counts each round of supplementary search operations triggered. After each supplementary search is completed, the current iteration count is incremented by 1, and the total number of supplementary searches executed is recorded in real time to ensure that the iteration count can be accurately tracked and verified in real time. If the comparison result shows that the number of supplementary search iterations exceeds the maximum number of iterations, regardless of whether the currently obtained evidence has reached the ideal level of sufficiency, the fourth agent node immediately terminates all supplementary search processes, organizes and consolidates all the multi-source evidence that has been retrieved, and forcibly outputs this evidence as multi-source evidence to be fused, directly entering the subsequent evidence fusion stage.

[0080] This embodiment sets a maximum number of iterations for supplementary retrieval, and forcibly terminates the retrieval and outputs existing evidence after the limit is exceeded, thus avoiding infinite loops of supplementary retrieval and ensuring the stable and efficient progress of the question-and-answer process.

[0081] Furthermore, in some embodiments, step S60, "receiving multi-source evidence to be fused through the fifth intelligent agent node, fusing the retrieved multi-source evidence, generating a structured answer, and outputting the structured answer and its cited sources," may specifically include: S601. Integrate the search results from different knowledge sub-bases according to logical relationships; Specifically, for step S601, the fifth intelligent agent node first performs text parsing and feature extraction on all received multi-source evidence. Through keyword matching, entity recognition, and source tag matching, the evidence is categorized into layers from two dimensions. By knowledge source, the evidence is mapped to its corresponding knowledge base type, distinguishing between enterprise disclosure evidence from the enterprise ESG report knowledge sub-base, indicator specification evidence from the ESG indicator semantic enhancement knowledge sub-base, and regulatory clause evidence from the ESG regulations and knowledge graph knowledge sub-base. By evidence type, based on content attributes, evidence is divided into factual data evidence (specific values, implementation measures), compliance rule evidence (regulatory requirements, binding clauses), semantic interpretation evidence (indicator definitions, connotation explanations), and comprehensive supporting evidence (supplementary explanations, related information).

[0082] Each piece of evidence is labeled with a corresponding category tag, and the chaotic and disordered multi-source evidence is organized into a well-defined and orderly set of evidence, avoiding the mixing of different types of evidence and providing a standardized basis for subsequent integration.

[0083] For example, the evidence obtained includes three items: "Company A reduced carbon emissions by 8,000 tons in 2024", "Company greenhouse gas emission accounting method", and "Carbon emission limit standard for key polluting units". According to the source, it is divided into enterprise report evidence, indicator semantic evidence and regulatory clause evidence; according to the type, it is divided into factual data evidence, semantic interpretation evidence and compliance rule evidence, and the classification and sorting are completed.

[0084] S602. If there are conflicts in the information of different knowledge sub-bases, the conflicting content and information from each source shall be explicitly marked; Specifically, for step S602, the core text and key data of the evidence within the group are compared to identify completely duplicated or highly similar content. The most complete piece of information is retained, and redundant statements are deleted. If there are contradictions in the data or statements of different pieces of evidence within the group, the valid information is determined according to the authority priority rule, with the priority being: legal clause evidence > indicator and standard evidence > corporate disclosure evidence. The content of high-authority evidence is retained, and the biased information of low-authority evidence is corrected. Complementary evidence is spliced ​​together to sort out the supporting, progressive, and explanatory relationships between the evidence. Scattered information fragments are integrated into logically coherent and complete information paragraphs to ensure that the integrated information is free of contradictions, omissions, and redundancy.

[0085] For example, in corporate reporting evidence, one statement reads "Company A reduced carbon emissions by 8,000 tons in 2024," while another reads "Company A reduced carbon emissions by approximately 8,000 tons in 2024." Duplicate content is removed to retain accurate data. In regulatory evidence, if there are slight differences between two limit standards, the latest published regulatory clauses shall prevail, and conflicts shall be resolved to integrate them into a complete compliance requirement text.

[0086] S603. Sort the fused information according to the preset priority rules and generate a structured answer. The structured answer shall contain at least three levels of structure: core conclusion, supporting evidence and detailed explanation. The preset priority rules are: regulatory information takes precedence over corporate environmental and social governance report information, corporate environmental and social governance report information takes precedence over market opinion information, and market opinion information takes precedence over online search information. Specifically, for step S603, based on the complete evidence that has been integrated and combined with the type of user question, a standardized paradigm is used to build the answer structure.

[0087] For compliance judgment questions, the structure of conclusion first + compliance basis + factual verification is adopted; For factual questions about enterprises, a structure of core data plus supplementary explanations should be used. For comprehensive comparison questions, the structure of comparison dimensions + data for each subject + summary of differences is adopted; The integrated evidence is then placed into the corresponding structural framework, and the content is organized in the form of paragraphs and key points to avoid cluttered text and to make the answers clear, logical, and in line with professional reading habits.

[0088] For example, regarding the question of "whether Company A's carbon emissions in 2024 are compliant", the following structured content is generated according to the paradigm: First, the conclusion is given that "Company A's carbon emissions in 2024 meet the relevant regulatory limits", then the basis for compliance is listed as "the carbon emission limit for key polluting units is 10,000 tons per year", and finally, the factual verification is added: "Company A's actual carbon emissions in 2024 were 8,000 tons, which is lower than the limit standard".

[0089] For key information such as core data, legal clauses, and indicator definitions in structured answers, reverse tracing is performed using unique evidence identifiers to accurately locate the original source of each key piece of information. The tracing content includes knowledge sub-base type, original document number, clause number, paragraph index, etc. Then, the source information obtained from the tracing is matched one by one with the key information in the answer, and the source is marked with parentheses and end-of-text indexes to ensure that each key piece of content has a corresponding traceable basis, thereby improving the rigor and credibility of the answer.

[0090] The generated structured answer text is integrated with all the cited source information that has been marked, and packaged in the form of "main answer + citation source details" and output to the interactive interface simultaneously. The structured answer is used to intuitively display the answer content, and the citation source details are used to provide a complete reference path, presenting the final result of the entire question and answer.

[0091] This embodiment classifies, sorts, and integrates multi-source evidence to resolve conflicts, then generates a well-structured answer with clear hierarchy according to a standardized paradigm, and marks the source of key information. The final output is logically coherent, standardized, and complete, while also possessing authoritative traceability, greatly improving the professionalism and credibility of ESG question-and-answer results.

[0092] Furthermore, in some embodiments, the method further includes: S604. Explicitly express any uncertainty in the answers, and / or label the credibility level of online search results.

[0093] Specifically, for step S604, after generating the structured answer, the answer text is subjected to sentence-by-sentence semantic verification and information validation. The content with uncertain characteristics is identified by preset rules. This type of content mainly includes: the search evidence has slight conflicts and cannot be completely concluded; the key data is only supported by a single source; some information has not been confirmed by authoritative laws or corporate reports; the expression has speculative or unverified characteristics; the statistical scope of the data is unclear, etc. The above content is marked as uncertain information to be processed one by one.

[0094] For marked uncertain content, use intuitive and clear text descriptions for explicit labeling, without concealing the information's unverifiable attributes. Specifically, add a prompt prefix before the corresponding content and add supplementary explanations at the end of the sentence to clearly indicate that the content has speculative, insufficient evidence, or unverifiable attributes, allowing users to clearly distinguish between certain conclusions and uncertain information.

[0095] For supplementary information obtained through online search channels cited in the answers, the credibility level of each online search result is comprehensively determined based on dimensions such as the authority of the information source, the qualifications of the publishing entity, and the verifiability of the content. Credibility levels are categorized into three tiers based on authority: high credibility corresponds to official regulatory platforms, authoritative industry institutions, and official corporate release channels; medium credibility corresponds to reputable industry media and professional research institutions; and low credibility corresponds to ordinary self-media, informal information platforms, and anonymous content.

[0096] The completed credibility level will be clearly labeled and attached next to the corresponding online search results, or simultaneously marked in the list of cited sources in the answer, to intuitively show the reliability of the online information and make it easier for users to intuitively judge the reference value of the information.

[0097] This embodiment clearly informs users of the reliability of information by explicitly prompting uncertain content in the answers and classifying and labeling the credibility level of online search information, thus avoiding misleading information and improving the rigor and transparency of the question-and-answer results.

[0098] Furthermore, in some embodiments, the multi-source environmental and social governance knowledge base also includes a market rating and network supplementary sub-base, which is used to store market environmental and social governance rating data and timely network information. The market rating and network supplementary sub-base serves as a supplementary information source during retrieval.

[0099] Specifically, this embodiment adds an independent market rating and online supplementary sub-repository to the existing multi-source Environmental and Social Governance (ESG) knowledge base system. This sub-repository, as an extension of the knowledge base, together with the original knowledge modules, forms a complete multi-source ESG knowledge base. It is specifically designed to carry supplementary ESG information not covered by the regular knowledge sub-repository, and does not replace the core storage function of the original sub-repository; it only serves as an extension module to improve the knowledge system. ESG-related rating information released by various professional rating agencies and industry assessment organizations is collected, organized, and standardized. The organized market ESG rating data is stored in this supplementary sub-repository. The stored content includes professional market evaluation data such as enterprise ESG rating levels, rating scoring details, rating dimension divisions, industry rating averages, and rating adjustment explanations, forming a complete set of market rating data.

[0100] By crawling and filtering publicly released ESG-related timely information online, and removing invalid and false information, the real-time online information is stored in this supplementary sub-database. This type of information mainly includes the latest industry policy updates, temporary ESG-related announcements from enterprises, sudden environmental and governance events, and cutting-edge ESG practice cases, which are dynamic information that is updated quickly and cannot be included in regular reports or regulatory databases in a timely manner.

[0101] During the ESG knowledge retrieval process, the market rating and online supplementary sub-database is not used as the primary core information source. It is only invoked after the core knowledge retrieval is completed, or when it is necessary to supplement the market evaluation perspective or real-time dynamic information. The data in the sub-database is used to supplement the retrieval results, thereby improving the dimensionality and timeliness of the retrieval information.

[0102] This embodiment adds a supplementary sub-base to the multi-source ESG knowledge base and stores market ratings and timely online information, while using it as a supplementary source for retrieval. On the basis of retaining the core knowledge system, it expands the coverage and real-time nature of ESG knowledge, making the knowledge information more comprehensive.

[0103] To facilitate understanding of the intelligent question-answering method based on multi-source ESG knowledge base and agent reasoning provided in this embodiment, such as... Figure 3 As shown, this embodiment also provides another implementation of an intelligent question-answering method based on multi-source ESG knowledge bases and agent reasoning. It constructs an agent reasoning process based on LangGraph, introduces DeepResearch concepts, and combines a multi-source ESG knowledge base to implement question answering, including the following steps: S1: Construct a multi-source ESG knowledge base, which includes a two-layer vector library of ESG reports, a semantic enhancement vector library of indicators, a regulatory and knowledge graph library, and a market rating and network supplementary library. The construction of a multi-source ESG knowledge base is the foundation for achieving accurate question answering. The specific construction methods of each sub-base are as follows: (1) ESG report dual-layer vector library: The dual-vector strategy of decoupling Summary / Content is adopted. For enterprise ESG reports (PDF / text), the text content is first extracted by PDF parsing tool, and then the Summary of each chapter of the report is generated by using BM25 algorithm combined with qwen-max large language model. The Summary content is vectorized by jina-embeddings-v3 model (vector dimension 1024) to build a Summary vector library for quick retrieval and location of core information. At the same time, the original text of ESG report is divided into paragraphs, and each paragraph is vectorized by Qwen3-Embedding-4B model (vector dimension 2560) to build a Content vector library to provide accurate factual and evidence support. This design avoids the semantic compression problem caused by a single vector carrying high-level semantics and high-density factual information at the same time.

[0104] (2) Indicator Semantic Enhancement Vector Library: To address the problem that the original ESG indicator names have extremely weak semantics and are difficult to directly participate in semantic retrieval, the ESG indicator names and corresponding rating points from mainstream standards such as ISO26000 and GRI are collected as input. Natural language definitions (including indicator connotation, calculation method, disclosure requirements, etc.) of each indicator are generated through LLM. Then, the combined text of "indicator name + rating points + definition" is vectorized through the Qwen3-Embedding-4B model (vector dimension 2560) to construct an indicator semantic enhancement vector library. This library serves as a semantic bridge to achieve semantic alignment between ESG reports, market ratings, and regulatory clauses. (3) Regulatory and Knowledge Graph Library: Collect ESG regulations, industry standards and self-built ESG knowledge graphs (including indicators, enterprises, industries, constraints, etc.) of major economies around the world. Adopt a large block strategy that prioritizes logical integrity and divide the library into blocks according to "regulatory clauses / knowledge graph logical units" (such as a single regulatory clause, a set of related indicator relationships, etc.). Then, use the Sentence-BERT model to vectorize the data and build a regulatory and knowledge graph library. At the same time, record the structural information of each block in the library (such as regulatory number, knowledge graph node relationship, etc.) for structural constraints in the subsequent recall stage and control the noise risk brought by large blocks. (4) Market rating and online supplementary library: integrate ESG rating data (structured) from mainstream institutions such as MSCI and FTSE Russell to build a market rating sub-library; access the API interface of Bing Academic and authoritative industry websites to obtain timely ESG-related information and build an online supplementary library; both types of data are standardized and stored to supplement market views and timely information. In addition, this step categorizes the data sources by knowledge type, as shown in the table below. This categorization directly affects the selection of recall strategies in the subsequent Planning phase. Factual knowledge typically comes from ESG reports and indicator databases, and its main purpose is to provide factual evidence for corporate ESG disclosures; rule-based knowledge typically comes from regulations and guidelines in law and knowledge graph databases, and mainly supports ESG compliance judgments; relational knowledge typically comes from ESG knowledge graphs, which can realize associative reasoning and comprehensive analysis; and time-sensitive knowledge typically comes from supplementary online sub-databases, which are used to supplement ESG-related information that is highly time-sensitive.

[0105] S2: Receive ESG-related questions from users, determine the complexity of the question through the Supervisor node of the intelligent agent. If it is a low-complexity question, directly enter the Simple Answer node to generate a quick response; if it is a high-complexity question, enter the Plan node. Specifically, the Supervisor node uses LLM (such as DeepSeek-V3.1) to determine the complexity of user questions, with preset judgment criteria: low-complexity questions are those that do not require multi-source knowledge collaboration and only require a single fact / concept answer (such as "What are Scope1 emissions?"); high-complexity questions are those that require multi-source knowledge integration, compliance judgment, and comprehensive analysis (such as "Does a company's carbon emission reduction measures in 2023 comply with EU CSRD regulations?"). If it is a low-complexity question, it directly enters the Simple Answer node, calling relevant data from the indicator semantic enhancement vector library or the regulatory and knowledge graph library to generate a quick response; if it is a high-complexity question, it enters the Plan node.

[0106] S3: The Plan node breaks down high-complexity problems, clarifies the problem type and the required knowledge type, and formulates a recall strategy based on the knowledge type and problem type; Specifically, the Plan node first uses LLM to break down highly complex problems, clarifying core requirements (such as compliance judgment, comprehensive comparison, etc.), the types of knowledge required (such as rule-based, factual, relational, etc.), and key elements such as the companies / industries / indicators involved. Then, based on the problem type, a targeted recall strategy is formulated, as follows: ① Concept / definition questions (e.g., "What are the core requirements of the GRI303 indicator?"): Prioritize recalling Definition data from the indicator semantic enhancement vector library and corresponding criterion data from the regulatory and knowledge graph library to ensure the accuracy and authority of the definition; ② For factual questions about enterprises (such as "What was the water consumption of a certain enterprise in 2023?"): Extract specific data from the relevant paragraphs in the enterprise's 2023 ESG report by retrieving the relevant sections on "water consumption" from the ESG report vector library; ③ Compliance judgment questions (such as "Does a company's employee health protection measures comply with the ISO26000 standard?"): The main data is based on the ISO26000 standard clauses in the regulations and knowledge graph database (to clarify compliance requirements), combined with the indicator semantic enhancement vector library to align the relevant indicators of "employee health protection" (to unify semantics), and then the specific measures of the company are extracted through the ESG report content vector library (to verify the actual disclosure). ④ Market perspective questions (e.g., "Why did MSCI give a certain company an ESG rating of BB?"): Retrieve core disclosure information from MSCI rating report data and ESG report vector library in the market rating and online supplementary database, supplemented by analysis articles from authoritative industry websites with low weight (supplementing market perspectives). ⑤ Comprehensive / comparative analysis questions (such as "comparing the carbon emission reduction performance of companies A and B in 2023"): Recall the results by integrating the "carbon emission reduction" indicator Definition (unified evaluation standard) from the semantic enhancement vector library, the core carbon emission reduction data from the ESG report summary vector library of companies A and B (quickly obtain key information), and the carbon emission reduction rating data of the two companies from the market rating and online supplementary library (combined with market evaluation). Meanwhile, to optimize the recall effect, the Plan node also generates "recall suggestions" in natural language (such as "This issue is a compliance judgment, and regulations should take priority"), realizing semantic priority control of the recall source, and achieving soft priority adjustment without introducing numerical weights. S4: The Task Execute node performs concurrent multi-source knowledge retrieval from the multi-source ESG knowledge base according to the defined recall strategy to obtain the initial retrieval results; Specifically, the concurrent retrieval in this step involves multiple core details, as follows: First, a hybrid vector retrieval strategy combined with Reciprocal Rank Fusion (RRF) is adopted for the ESG report knowledge base. The hybrid retrieval results are fused according to the RRF algorithm to balance the retrieval weights of each information source. The multi-dimensional hybrid retrieval strategy combined with RRF includes three types of retrieval channels: keyword sparse vector retrieval based on the BM25 algorithm (calculated after word segmentation of the original report text to achieve precise keyword matching), dense vector retrieval based on the original report text (using Qwen3-Embedding-4B 2560-dimensional dense vectors to represent the semantics of the full text), and vector semantic retrieval based on the summary of the original text (jina-embeddings-v3 1024-dimensional vectors to represent the semantics of the summary). The results of the three types of retrieval channels are fused according to the RRF algorithm to balance the weights of each retrieval dimension, thereby improving the comprehensiveness and reliability of the retrieval results.

[0107] Second, we implemented query rewriting and optimization for multiple knowledge bases. By using LLM to generate multiple rewritten queries from different dimensions such as semantic synonymy, scenario adaptation, and element completion, we achieved multi-perspective semantic retrieval and improved the recall rate of key information. Third, the Rerank model is introduced for fine ranking. Based on the semantic relevance of the search results to the user's question, the authority of the source, the completeness of the content, and other dimensions, the preliminary search results are re-ranked to filter high-value information. Fourth, perform layered deduplication processing, deduplicating duplicate content within a single retrieval task, and comparing results across different retrieval tasks to eliminate duplicate information across tasks, ensuring the uniqueness and conciseness of the initial retrieval results.

[0108] Specifically, the Task Execute node, based on the recall strategy defined by the Plan node, calls the retrieval interfaces of each sub-library for concurrent retrieval: semantic similarity matching for the vector library, SQL queries for structured data (such as market rating data), and API interfaces for retrieval of the network supplementary sub-library; after the retrieval is completed, the results are initially filtered (results with relevance scores below the threshold are removed) by combining the above-mentioned hybrid retrieval, query rewriting, reranking and hierarchical deduplication process to form the initial retrieval results, which include the original data, relevance scores and source information of each knowledge source.

[0109] S5: The Reflection node performs evidence sufficiency assessment based on LLM, taking into account the user's original goal, initial solution approach, current results, and currently executed tasks, and assesses the evidence from three dimensions: coverage of key knowledge types, completeness of evidence, and alignment of results with goals. If the evidence is insufficient, it returns to the Task Execute node to trigger supplementary retrieval; if the information is sufficient, it proceeds to the Reply node. Specifically, this step is the core of ensuring the quality of question and answer. The Reflection node uses LLM for meta-level judgment, does not directly generate answers, but only judges "whether enough is known". Its input information includes: user's original goal, initial solution approach (i.e., the decomposition result of the Plan node), current results (i.e., the initial search results of the Task Execute node), and current tasks (i.e., the completed search operations). The core judgment dimensions of the Reflection node include: ① whether the key knowledge types required for the question have been covered (e.g., compliance judgments need to cover rule-based and factual knowledge); ② whether there are obvious missing or weak evidence (e.g., lack of specific data disclosed by the company, incomplete citation of regulatory clauses, etc.); ③ whether the current results deviate from the user's goal (e.g., if the user asks "carbon emission reduction effectiveness", the result only involves "carbon emission data"). To enhance the standardization of the judgment, this step also introduces semi-structured constraints, with a pre-set checklist: "Has the company's disclosed information been searched?" "Has authoritative regulations been consulted?" "Is there at least one piece of factual evidence?" The LLM self-assesses the checklist, balancing the flexibility and standardization of the judgment without the need for hard-coded rules. If the Reflection node determines that the evidence is "insufficient" (e.g., no specific disclosure data from the company was found, key regulatory clauses are missing, etc.), it returns to the Task Execute node to trigger a supplementary search (e.g., re-searching the company's ESG report content vector library, supplementing the search for relevant regulatory clauses, etc.); if it determines that the information is "sufficient", it pushes the initial search results to the Reply node. S6: The Reply node performs multi-source evidence fusion on the search results, marks conflict information, generates a structured answer according to preset priority rules, and outputs the cited sources.

[0110] Specifically, in this step, the Reply node first performs fusion processing on the initial search results: ① Multi-source evidence fusion: Integrates relevant information from different knowledge sources according to logical connections (e.g., compliance judgments first present regulatory requirements, then correspond to the company's actual measures); ② Conflict information annotation: If there are conflicts between information from different knowledge sources (e.g., the carbon emissions disclosed in the company's report are inconsistent with third-party data), the conflicting content and information from each source are explicitly annotated; ③ Structured generation: Generates answers according to the structure of "core conclusions - supporting evidence - detailed explanation," adapting to the reading needs of ESG professional analysis scenarios; ④ Citation source output: Clearly annotates the source of each piece of information (e.g., "Company's 2023 ESG Report P25," "MSCI 2023 ESG Rating Report," etc.). The integration process follows a pre-defined priority rule: regulatory information > corporate ESG report information > market opinion information > online search information; at the same time, uncertainties in the answers (such as "estimated based on industry average level because the company did not disclose specific data") are explicitly stated to improve the reliability of the answers.

[0111] Furthermore, regarding the scenario-based use of the Summary vector library, if the current issue involves strict compliance or critical judgment scenarios, the Reply node will combine the judgment results of the Reflection node to avoid relying solely on the information in the Summary vector library. If it finds a lack of original text evidence, it will automatically return to the Task Execute node to supplement the search of the Content vector library, or downgrade the expression in the answer (e.g., "This conclusion is based on the report summary; detailed evidence needs to be referred to in the original text PXX"). For online search results, its credibility level will be explicitly marked (e.g., "Source is an authoritative industry website, highly credible" or "Source is a general blog, for reference only"), and it will not be used as the sole basis for compliance judgment.

[0112] Furthermore, in step S1, the ESG report two-layer vector library is constructed using a Summary / Content decoupling two-vector strategy, specifically including: S111: Parse the enterprise ESG report (PDF / text), generate a report summary using the BM25 algorithm combined with a large language model (LLM), vectorize the summary content (vector dimension 1024), and build a summary vector library for quick retrieval and location of core information; S112: Extract paragraphs from the original ESG report as content, vectorize the original paragraphs (vector dimension 2560), and build a content vector library to locate factual details and provide evidence support.

[0113] Further, in step S1, the construction process of the indicator semantic enhancement vector library includes: S121: Collect ESG indicator names and corresponding rating criteria as input elements; S122: Generate natural language interpretations (Definitions) for each ESG indicator using LLM. S123: Vectorize indicator names, rating points, and natural language explanations (vector dimension 2560) to build an indicator semantic enhancement vector library for alignment with ESG reports, market ratings, and regulatory provisions.

[0114] Furthermore, in step S1, the regulations and knowledge graph library are constructed using a large-block strategy that prioritizes logical integrity: ESG regulations, guidelines, and knowledge graphs (including indicators, relationships, and constraints) are divided into blocks according to "clauses / logical units" to avoid excessive fine-grained segmentation that could damage normative semantics and logical integrity, and structural constraints are introduced in the subsequent recall stage to control noise risks.

[0115] Furthermore, in step S3, the problem type-driven recall strategy is specifically as follows: If the question is a concept / definition type, prioritize recalling regulatory vector data from regulations and knowledge graph bases; If the question pertains to corporate facts, relevant information is queried in parallel using ESG reports and the extracted indicator database; If the question is a compliance judgment, the main focus is on the regulatory vector data in the regulations and knowledge graph library, combined with the semantic enhancement vector library of indicators for semantic alignment, and then the actual disclosure of the enterprise is verified through the ESG report vector library. If the question pertains to market opinions, recall market rating data and ESG report vector library data from the market rating and online supplementary databases, supplemented by low-weight online search data. If the problem is a comprehensive / comparative analysis type, recall data by integrating the semantic enhancement vector library of integrated indicators, the multi-enterprise ESG report vector library, and the rating data in the market rating and online supplementary library.

[0116] Furthermore, in step S5, the evidence sufficiency judgment of the Reflection node introduces semi-structured constraints. It uses an LLM (Limited Language Management) to self-assess a preset checklist, which includes: whether enterprise disclosure information has been retrieved, whether authoritative regulations have been referenced, and whether at least one piece of factual evidence exists. This improves the standardization of the judgment without requiring hard-coded rules. Additionally, to prevent the system from entering an infinite loop, a maximum number of iterations is set. When the maximum number of iterations is exceeded, the Reflection node will output "End query".

[0117] Furthermore, in step S6, the preset priority rule is: regulatory information > corporate ESG report information > market opinion information > web search information; at the same time, the uncertain content in the answer is explicitly expressed.

[0118] In a specific embodiment, this embodiment also provides an ESG intelligent question answering system that integrates a multi-source ESG knowledge base and is based on an agent reasoning process, for implementing the above-mentioned ESG intelligent question answering method. The system includes: a multi-source ESG knowledge base construction module, an agent reasoning module, and a result output module. The multi-source ESG knowledge base construction module is used to build and maintain an ESG report vector library, an indicator semantic enhancement vector library, a regulatory and knowledge graph library, and a market rating and network supplement library, so as to realize the integration and standardized storage of multi-source heterogeneous ESG knowledge. The agent reasoning module is built on LangGraph and includes Supervisor, Simple Answer, Plan, Task Execute, and Reflection nodes; wherein: Supervisor node: Used to determine the complexity of the user's question and decide whether to proceed to a complex reasoning process; Simple Answer node: Used for quick responses to low-complexity questions; Plan node: Used to break down highly complex problems, clarify knowledge requirements, and formulate problem type-driven recall strategies; Task Execute node: Used to concurrently retrieve knowledge from multi-source ESG knowledge bases according to the recall strategy and obtain initial results; Reflection node: Used to determine the sufficiency of evidence based on LLM, triggering supplementary retrieval or allowing the results to be output to the result output module; The result output module, corresponding to the Reply node in the method, is used to realize multi-source evidence fusion, conflict information annotation, structured response generation, and source citation output.

[0119] Furthermore, the multi-source ESG knowledge base construction module also includes a knowledge base optimization unit, which is used to: control the semantic priority of the recall source, generate recall suggestions in natural language form during the Plan stage; and optimize the proportion of different information sources according to different scenarios when multi-source information is fused. For example, in compliance scenarios, the credibility of search results is marked and used as a reference for compliance judgment.

[0120] Furthermore, the data sources in the multi-source ESG knowledge base include corporate ESG reports (PDF / text), corporate ESG indicator extraction results (structured), market ESG rating data, ESG regulations and guidelines, ESG knowledge graphs, and web search results (supplementary and timely). Each data source is classified into factual, overview, structured semantic, rule-based, relational, and timely types to support the selection of recall strategies in the Plan phase.

[0121] like Figure 4 As shown, this embodiment provides an ESG intelligent question answering system that integrates a multi-source ESG knowledge base and is based on an agent reasoning process, for implementing the above-mentioned ESG intelligent question answering method. The system includes: a multi-source ESG knowledge base construction module, an agent reasoning module, and a result output module. 1. Multi-source ESG knowledge base construction module, used to build and maintain a two-layer vector library for ESG reports, a semantic enhancement vector library for indicators, a regulatory and knowledge graph library, and a market rating and network supplementary library, to achieve the integration and standardized storage of multi-source heterogeneous ESG knowledge; this module includes a data parsing unit, a vector generation unit, a library management unit, and a knowledge base optimization unit; Data parsing unit: used to parse corporate ESG reports (PDF / text), extract ESG indicators, organize regulatory and knowledge graph data, clean market rating data and online supplementary data, and output standardized text / structured data; Vector generation unit: Equipped with models such as jina-embeddings-v3 and Qwen3-Embedding-4B, it is used to vectorize various types of text data and generate vector data of different dimensions; Library Management Unit: Used to store vector data, structured data and source information of each sub-library, and provides retrieval interface and data update interface; Knowledge base optimization unit: used to implement semantic priority control of recall sources (generating natural language recall suggestions), scenario-based use of ESG report vector library (detecting the integrity of original text evidence), and credibility management of search results (labeling credibility level). 2. The agent reasoning module, built on Lang Graph, includes Supervisor, Simple Answer, Plan, Task Execute, and Reflection nodes; the functions of each node are as follows: Supervisor node: Uses LLM to determine the complexity of the user's problem and decides whether to proceed to a complex reasoning process; Simple Answer node: Calls on basic data from multiple ESG knowledge bases to quickly respond to low-complexity questions; Plan node: Decompose high-complexity problems, clarify knowledge requirements, formulate problem type-driven recall strategies, and generate semantic priority recall suggestions; Task Execute node: Calls the retrieval interface of the multi-source ESG knowledge base, performs concurrent retrieval according to the recall strategy, obtains initial retrieval results and performs preliminary filtering; Reflection node: Based on LLM and semi-structured checklist, it judges the sufficiency of evidence for the initial search results, triggers supplementary search or allows the results to be output. 3. The results output module, corresponding to the Reply node, is used to realize multi-source evidence fusion, conflict information annotation, structured answer generation, and citation source output; at the same time, combined with the feedback from the knowledge base optimization unit, it realizes the scenario-based adaptation of the Summary vector library and the credibility annotation of network search results. In this embodiment, the data sources in the multi-source ESG knowledge base include corporate ESG reports (PDF / text), corporate ESG indicator extraction results (structured), market ESG rating data, ESG regulations and guidelines, ESG knowledge graphs, and web search results (supplementary and timely). Each data source is classified into factual, overview, structured semantic, rule-based, relational, and timely types to support the selection of recall strategies in the Plan phase.

[0122] In summary, compared with existing technologies, the intelligent question-answering method based on multi-source ESG knowledge bases and agent reasoning provided in this embodiment achieves the classification, integration, and efficient collaboration of heterogeneous ESG knowledge through the refined construction of multi-source ESG knowledge bases. In particular, the design of the two-layer vector library for ESG reports and the semantic enhancement vector library for indicators solves key problems such as semantic compression of single vector libraries and weak semantics of indicators. The agent reasoning process based on LangGraph, combined with a question type-driven recall strategy, achieves accurate adaptation to different ESG question-answering scenarios and improves the accuracy of answers. The evidence sufficiency judgment mechanism of the Reflection node introduces semi-structured constraints, which can automatically judge whether "information is sufficient" to avoid analysis bias caused by insufficient or redundant information, while ensuring the interpretability of the answers. The priority rules and explicit expression of uncertainty in the result fusion stage further improve the professionalism and reliability of the answers and adapt to the needs of professional scenarios such as ESG compliance analysis.

[0123] To facilitate better implementation of the intelligent question answering method based on multi-source ESG knowledge base and agent reasoning in the embodiments of this application, this application also provides an intelligent question answering device based on multi-source ESG knowledge base and agent reasoning, which is based on the aforementioned intelligent question answering method based on multi-source ESG knowledge base and agent reasoning. The meanings of the terms used are the same as in the aforementioned intelligent question answering method based on multi-source ESG knowledge base and agent reasoning; specific implementation details can be found in the descriptions in the method embodiments.

[0124] Please see Figure 5 , Figure 5This is a schematic diagram of the structure of an intelligent question-answering device based on a multi-source ESG knowledge base and agent reasoning provided in an embodiment of this application. Specifically, the intelligent question-answering device may include a knowledge base construction module 201, a first reasoning module 202, a second reasoning module 203, a third reasoning module 204, a fourth reasoning module 205, and a fifth reasoning module 206, as detailed below: The knowledge base construction module 201 is used to construct a multi-source environmental and social governance knowledge base. The multi-source environmental and social governance knowledge base includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. The first reasoning module 202 is used to receive environmental and social governance questions input by the user, and to determine the complexity of the environmental and social governance questions through the first intelligent agent node. If the complexity is low, a quick answer is generated and output; if the complexity is high, the question information to be broken down is output. The second reasoning module 203 is used to receive the problem information to be decomposed through the second intelligent agent node, decompose the highly complex environmental and social governance problem, determine the knowledge type required to answer the environmental and social governance problem, generate a recall strategy based on the knowledge type, and output the recall strategy. The third reasoning module 204 is used to receive the recall strategy through the third intelligent agent node, and according to the recall strategy, concurrently retrieve knowledge related to environmental and social governance issues from the multi-source environmental and social governance knowledge base to obtain initial search results. The fourth reasoning module 205 is used to make a sufficiency judgment of evidence based on the initial search results through the fourth intelligent agent node. If the evidence is not sufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is sufficient, the multi-source evidence to be fused is output. The fifth reasoning module 206 is used to receive multi-source evidence to be fused through the fifth intelligent agent node, fuse the retrieved multi-source evidence, generate a structured answer, and output the structured answer and its cited sources.

[0125] For specific limitations regarding the intelligent question-answering device based on multi-source ESG knowledge bases and agent reasoning, please refer to the limitations of the intelligent question-answering method based on multi-source ESG knowledge bases and agent reasoning mentioned above, which will not be repeated here. Each module in the aforementioned intelligent question-answering device based on multi-source ESG knowledge bases and agent reasoning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0126] The intelligent question-answering device based on a multi-source ESG knowledge base and agent reasoning provided in this embodiment constructs a multi-source ESG knowledge base, and multiple agent nodes collaboratively complete question processing, knowledge retrieval, evidence verification, and result fusion to achieve accurate and efficient structured intelligent question answering in the ESG field. Finally, it forms rigorous and standardized ESG intelligent question-answering results, thereby improving the completeness and reliability of the question-answering results and being able to stably adapt to the answering needs of various ESG issues.

[0127] Furthermore, embodiments of this application also provide an electronic device, such as... Figure 6 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically: The electronic device may include components such as a processor 301 with one or more processing cores, a memory 302 with one or more computer-readable storage media, a power supply 303, and an input unit 304. Those skilled in the art will understand that... Figure 6 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: The processor 301 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory 302, and by calling data stored in the memory 302, thereby providing overall monitoring of the electronic device. Optionally, the processor 301 may include one or more processing cores; preferably, the processor 301 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 301.

[0128] The memory 302 can be used to store software programs and modules. The processor 301 executes various functional applications and intelligent question-answering methods based on multi-source ESG knowledge bases and agent reasoning by running the software programs and modules stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 302 may also include a memory controller to provide the processor 301 with access to the memory 302.

[0129] The electronic device also includes a power supply 303 that supplies power to various components. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 303 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0130] The electronic device may also include an input unit 304, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0131] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 301 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 302 according to the following instructions, and the processor 301 runs the applications stored in the memory 302 to realize various functions, as follows: The system constructs a multi-source environmental and social governance knowledge base. It receives user-input environmental and social governance questions and, through a first intelligent agent node, determines the complexity of the question. If the complexity is low, a quick answer is generated and output; if it is high, information about the question to be broken down is output. A second intelligent agent node receives the information about the question to be broken down, decomposes high-complexity environmental and social governance questions to determine the types of knowledge required to answer them, generates a recall strategy based on the knowledge types, and outputs the recall strategy. A third intelligent agent node receives the recall strategy and, based on the strategy, concurrently retrieves knowledge related to the environmental and social governance question from the multi-source environmental and social governance knowledge base to obtain initial search results. A fourth intelligent agent node performs an evidence sufficiency judgment based on the initial search results. If the evidence is insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node; if the evidence is sufficient, multi-source evidence to be fused is output. A fifth intelligent agent node receives the multi-source evidence to be fused, fuses the retrieved multi-source evidence, generates a structured answer, and outputs the structured answer and its cited sources.

[0132] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0133] This application embodiment constructs a multi-source ESG knowledge base, in which multiple intelligent agent nodes collaboratively complete problem processing, knowledge retrieval, evidence verification, and result fusion, achieving accurate and efficient structured intelligent question answering in the ESG field. Ultimately, it forms rigorous and standardized ESG intelligent question answering results, thereby improving the completeness and reliability of the question answering results and enabling stable adaptation to the answering needs of various ESG issues.

[0134] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0135] To this end, embodiments of this application provide a storage medium storing multiple instructions that can be loaded by a processor to execute steps in any of the intelligent question-answering methods based on multi-source ESG knowledge bases and agent reasoning provided in embodiments of this application. For example, the instructions can execute the following steps: The system constructs a multi-source environmental and social governance knowledge base. It receives user-input environmental and social governance questions and, through a first intelligent agent node, determines the complexity of the question. If the complexity is low, a quick answer is generated and output; if it is high, information about the question to be broken down is output. A second intelligent agent node receives the information about the question to be broken down, decomposes high-complexity environmental and social governance questions to determine the types of knowledge required to answer them, generates a recall strategy based on the knowledge types, and outputs the recall strategy. A third intelligent agent node receives the recall strategy and, based on the strategy, concurrently retrieves knowledge related to the environmental and social governance question from the multi-source environmental and social governance knowledge base to obtain initial search results. A fourth intelligent agent node performs an evidence sufficiency judgment based on the initial search results. If the evidence is insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node; if the evidence is sufficient, multi-source evidence to be fused is output. A fifth intelligent agent node receives the multi-source evidence to be fused, fuses the retrieved multi-source evidence, generates a structured answer, and outputs the structured answer and its cited sources.

[0136] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0137] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0138] Since the instructions stored in the storage medium can execute the steps in any of the intelligent question-answering methods based on multi-source ESG knowledge bases and agent reasoning provided in the embodiments of this application, the beneficial effects that any of the intelligent question-answering methods based on multi-source ESG knowledge bases and agent reasoning provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0139] The above provides a detailed description of an intelligent question-answering method and apparatus based on a multi-source ESG knowledge base and agent reasoning, as provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning, characterized in that, include: Construct a multi-source environmental and social governance knowledge base, which includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. The system receives environmental and social governance questions input by users, and determines the complexity of the environmental and social governance questions through the first intelligent agent node. If the complexity is low, a quick answer is generated and output; if the complexity is high, the problem information to be broken down is output. The second intelligent agent node receives the problem information to be decomposed, decomposes the highly complex environmental and social governance problem, determines the type of knowledge required to answer the environmental and social governance problem, generates a recall strategy based on the knowledge type, and outputs the recall strategy. The third intelligent agent node receives the recall strategy and, based on the recall strategy, concurrently retrieves knowledge related to the environmental and social governance problem from the multi-source environmental and social governance knowledge base to obtain initial search results. The fourth intelligent agent node performs an evidence sufficiency judgment based on the initial search results. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, multi-source evidence to be fused is output. The fifth intelligent agent node receives the multi-source evidence to be fused, fuses the retrieved multi-source evidence, generates a structured answer, and outputs the structured answer and its cited sources.

2. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning as described in claim 1, characterized in that, The construction methods of the first knowledge sub-base include: Obtain the corporate environmental and social governance report text, parse the report text, and divide it into multiple paragraphs; Based on a preset language model, generate summary information for each paragraph, convert the summary information into a first semantic vector, and construct a summary vector sub-library; Each original paragraph of the report text is converted into a second semantic vector to construct a content vector sub-library; wherein the first semantic vector and the second semantic vector have different dimensions, the summary vector sub-library is used to quickly locate core information, and the content vector sub-library is used to provide factual details and evidence.

3. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning as described in claim 1, characterized in that, The construction methods of the second knowledge sub-base include: Collect the names of environmental and social governance indicators and their corresponding rating criteria; For each environmental and social governance indicator, a natural language explanation text is generated using a language model. The natural language explanation text includes at least one of the indicator's connotation, calculation method, or disclosure requirements. The names of the environmental and social governance indicators, the rating criteria, and the natural language explanation text are combined and converted into semantic vectors to construct a semantically enhanced vector sub-library for semantic alignment between different environmental and social governance knowledge sources.

4. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning as described in claim 1, characterized in that, The construction methods of the third knowledge sub-base include: Collect environmental and social governance (ESG) regulations and / or ESG knowledge graph data, wherein the knowledge graph data includes indicators, enterprises, industries and constraint relationships; Following a block-segmentation strategy prioritizing logical integrity, the legal provisions or knowledge graph data are divided into blocks according to logical units, so that each block maintains complete normative semantics. Each block is converted into a semantic vector, and the structural information of each block is recorded to construct a regulatory and knowledge graph vector sub-library, which is used to apply structural constraints during the recall phase.

5. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning according to claim 1, characterized in that, The process involves receiving the problem information to be decomposed through a second intelligent agent node, breaking down the highly complex environmental and social governance problem to determine the type of knowledge required to answer it, generating a recall strategy based on the knowledge type, and outputting the recall strategy, including: Identify the core needs, required knowledge types, and key elements involved in the environmental and social governance issues, including enterprises, industries, or indicators. The question types are determined based on the core requirements, including concept definition, business fact, compliance judgment, market opinion, or comprehensive comparative analysis. Based on the question type, a corresponding recall strategy is generated, wherein the recall strategies corresponding to different question types prioritize recalling knowledge from different knowledge sub-bases.

6. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning as described in claim 5, characterized in that, The generation of a corresponding recall strategy based on the problem type includes: If the question type is a concept definition type, data will be retrieved from the second or third knowledge sub-base first. If the question type is a business fact, data will be retrieved from the first knowledge sub-base first. If the problem type is a compliance judgment, data is retrieved first from the third knowledge sub-base, and semantic alignment is performed in conjunction with the second knowledge sub-base. Then, the enterprise's disclosure is verified through the first knowledge sub-base. If the question type is market opinion, data will be retrieved first from the market rating and online supplementary sub-library, supplemented by data from the first knowledge sub-library; If the problem type is a comprehensive comparative analysis, then the second knowledge sub-base, the first knowledge sub-bases of multiple enterprises, and market rating data are integrated for recall.

7. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning according to claim 1, characterized in that, The fourth agent node performs an evidence sufficiency assessment based on the initial search results. If the evidence is deemed insufficient, a supplementary search is triggered, and the supplementary search instruction is returned to the third agent node. If the evidence is deemed sufficient, multi-source evidence to be fused is output, including: Obtain the user's original goal, the decomposition results output by the second intelligent agent node, the initial search results, and the search operation information that has been executed; Based on preset sufficiency evaluation dimensions, the initial search results are judged for sufficiency of evidence. The sufficiency evaluation dimensions include the coverage of key knowledge types, the completeness of evidence, and the degree of fit between the search results and the user's goals. If the evidence is deemed insufficient, a supplementary search instruction is generated and returned to the third agent node to trigger a supplementary search. If the evidence is deemed sufficient, then output the multi-source evidence to be fused.

8. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning according to claim 7, characterized in that, The method of judging the sufficiency of evidence of the initial search results based on the preset sufficiency evaluation dimensions also includes: Semi-structured constraints are introduced, and a language model is used to perform a self-assessment of a preset checklist. The checklist includes whether the company's disclosed information has been retrieved, whether authoritative regulations have been referenced, and whether at least one piece of factual evidence exists.

9. The intelligent question-answering method based on a multi-source ESG knowledge base and agent reasoning according to claim 1, characterized in that, The process of receiving the multi-source evidence to be fused through the fifth intelligent agent node, fusing the retrieved multi-source evidence, generating a structured answer, and outputting the structured answer and its cited sources includes: The search results from different knowledge sub-bases are integrated according to logical relationships; If there are conflicts in the information of different knowledge sub-bases, the conflicting content and information from each source will be explicitly marked. The merged information is sorted according to a preset priority rule, and a structured answer is generated. The structured answer contains at least three levels: core conclusion, supporting evidence, and detailed explanation. The preset priority rule is that regulatory information takes precedence over corporate environmental and social governance report information, corporate environmental and social governance report information takes precedence over market opinion information, and market opinion information takes precedence over online search information.

10. An intelligent question-answering device based on a multi-source ESG knowledge base and agent reasoning, characterized in that, include: The knowledge base construction module is used to construct a multi-source environmental and social governance knowledge base, which includes at least one of the following: a first knowledge sub-base for storing enterprise environmental and social governance report data, a second knowledge sub-base for storing semantically enhanced data of environmental and social governance indicators, and a third knowledge sub-base for storing environmental and social governance laws and knowledge graph data. The first reasoning module is used to receive environmental and social governance questions input by the user, and to determine the complexity of the environmental and social governance questions through the first intelligent agent node. If the complexity is low, a quick answer is generated and output; if the complexity is high, the question information to be broken down is output. The second reasoning module is used to receive the problem information to be decomposed through the second intelligent agent node, decompose the highly complex environmental and social governance problem, determine the knowledge type required to answer the environmental and social governance problem, generate a recall strategy based on the knowledge type, and output the recall strategy. The third reasoning module is used to receive the recall strategy through the third intelligent agent node, and according to the recall strategy, concurrently retrieve knowledge related to the environmental and social governance problem from the multi-source environmental and social governance knowledge base to obtain initial search results; The fourth reasoning module is used to determine the sufficiency of evidence based on the initial search results through the fourth intelligent agent node. If the evidence is deemed insufficient, a supplementary search is triggered and the supplementary search instruction is returned to the third intelligent agent node. If the evidence is deemed sufficient, the multi-source evidence to be fused is output. The fifth reasoning module is used to receive the multi-source evidence to be fused through the fifth intelligent agent node, fuse the retrieved multi-source evidence, generate a structured answer, and output the structured answer and its cited sources.

Images