Question and answer method and system fusing multi-granularity knowledge graph and self-adaptive agent

Abstract

The invention relates to a question and answer method and system fusing a multi-granularity knowledge graph and a self-adaptive agent. The question and answer method comprises the steps of knowledge governance and extraction, knowledge generation and expansion, knowledge retrieval and management, and knowledge rule mining and self-adaptive retrieval. Through cooperation of the multi-granularity knowledge graph and the self-adaptive agent, questions of the user can be more comprehensively understood, and more relevant knowledge can be retrieved from the knowledge base, so that the question and answer accuracy is remarkably improved, and wrong answers are reduced. Meanwhile, the multi-granularity knowledge graph can support more complex knowledge application requirements such as reasoning and prediction, so that the application capability of the question and answer system is enhanced, and the question and answer system can process more complex question and answer tasks such as multi-round dialogues and multi-task question and answer.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and natural language processing, and in particular to a question-answering method and system integrating multi-granularity knowledge graphs and adaptive intelligent agents. Background Art

[0002] With the rapid development of artificial intelligence technology, question-answering systems have become one of the important applications in the field of natural language processing. Question-answering systems can understand natural language questions raised by users and give corresponding answers. They play an important role in various fields, including but not limited to intelligent customer service, information retrieval, educational assistance, medical consultation and other scenarios. The implementation methods of question-answering systems can be divided into four categories, namely traditional question-answering system methods, question-answering systems based on knowledge graphs, question-answering systems based on deep learning, and question-answering systems based on large language models.

[0003] Traditional question-answering systems are mainly based on the theories and methods of linguistics or statistical linguistics. The knowledge base-based method structures the information in the knowledge base, retrieves relevant knowledge based on the user's question, and generates answers. The template matching-based method matches the user's question with a pre-defined template and generates answers based on the information in the template. The slot-based method decomposes the user's question into multiple slots, each slot corresponds to a specific question type, and retrieves and generates answers based on the slot information.

[0004] The question-answering system based on the knowledge graph uses the knowledge graph to represent entities, concepts and relationships in the form of a graph structure, and then uses the rich information in the knowledge graph to perform semantic reasoning and knowledge question-answering. For example, the entities and relationships in the user's question can be extracted and then the matching paths can be found in the knowledge graph, and the matching paths are the corresponding answers.

[0005] The deep learning-based approach regards question answering as a sequence-to-sequence generation task, and often uses an encoder-decoder network structure, where the encoder is responsible for encoding the user's question into a vector representation, and the decoder generates the answer based on the context information, thereby achieving the effect of question answering. Usually, the specific network implementation of the encoder-decoder can be a recurrent neural network (RNN), a long short-term memory network (LSTM), a Transformer, etc. Transformer is a deep neural network model based on the self-attention mechanism, which can effectively capture the long-distance dependencies in the text, making the overall generated answer more accurate.

[0006] Question answering methods based on large language models have made significant progress in recent years and have gradually become the mainstream method in the field of question answering systems. The core idea of ​​this type of method is to regard question answering as a sequence-to-sequence generation task, and use deep learning models, especially Transformer models, to understand and generate natural language. Large language models are usually pre-trained on a large amount of corpus data to learn rich language knowledge and semantic representations, so that they can better handle various types of question answering tasks. When it is necessary to adapt to specific downstream question answering tasks, the pre-trained model can be fine-tuned, for example, it can be trained on corpus data in a specific field to better understand the knowledge in that field.

[0007] Although large language models are quite capable in question-answering, they are also prone to hallucinations (generating answers that appear reasonable but are actually wrong). At the same time, large language models have limited knowledge of specific fields and are not competent for question-answering tasks in professional fields. To solve these problems, existing technical solutions attempt to enhance the accuracy of question-answering systems with the help of external knowledge.

[0008] 1. Adopting the Retrieval-Augmented Generation (RAG) technology route, the external knowledge is constructed and stored in the vector database. The support points of the answer are found by comparing the vector similarity between the question and the knowledge, and then the large language model summarizes and generates the answer.

[0009] 2. Use knowledge graphs to enhance large model question answering. Information such as entities, relationships, and attributes in the knowledge graph is embedded into the vector space and provided as input to the large language model, which helps the model better understand the entities and relationships in the question and retrieve relevant knowledge from the knowledge base.

[0010] 3. Use agents to enhance large model answers. The large language model can communicate with other agents. For example, the large language model can communicate with an agent dedicated to retrieving information to obtain more accurate information and integrate it into the answer generation process. Multiple agents can also be used to work together to complete more complex question-answering tasks, such as multi-round dialogues and multi-task question-answering.

[0011] Accordingly, the solutions in the prior art have the following deficiencies:

[0012] 1. Text content does not exist in isolation, but is interrelated. Traditional RAG cannot capture important structured relational knowledge that cannot be presented by semantic similarity alone. For example, in a citation network that connects papers through citation relationships, traditional RAG methods focus on finding relevant papers based on queries, but ignore the important citation relationships between papers.

[0013] 2. Most existing knowledge graph embedding methods only focus on the semantic representation of entities and relationships, while ignoring other information such as entity attributes, resulting in limited semantic understanding capabilities of the model. The construction of knowledge graphs requires a large amount of manual annotation and expert knowledge, which is costly and difficult to expand to new fields. Most existing knowledge graph reasoning methods are based on simple rule reasoning, which is difficult to handle complex logical reasoning and has low efficiency.

[0014] 3. The design of the intelligent agent needs to consider multiple aspects such as dialogue strategy, knowledge base management, and reasoning mechanism, which is complex. The decision-making process of the intelligent agent is usually difficult to explain, and it is difficult to gain the trust of users in scenarios with complex logical reasoning. Summary of the invention

[0015] With the rapid development of large language models, more and more industries and fields have begun to use large language models for daily office assistance, or to act as intelligent customer service, or to assist in searching and writing materials. However, the hallucination phenomenon of large language models is very serious, and they often generate some answers that look correct but are actually wrong. On the other hand, although large language models use a large amount of corpus data in the pre-training stage, they know very little about the vertical fields in actual application scenarios. How to effectively apply the knowledge of vertical subdivisions to large language models, build a systematic and comprehensive question-and-answer system, and effectively improve the accuracy of their answers is a problem that needs to be solved urgently.

[0016] In view of this, the present invention provides a question-answering method integrating a multi-granularity knowledge graph and an adaptive agent, which comprises:

[0017] Step 1: Knowledge governance and extraction.

[0018] Step 1.1: Data access. Support the access of multi-source heterogeneous data, including text, images, audio, etc., and uniformly process them through multimodal intelligent semantic models to convert non-text data into text form.

[0019] Step 1.2: Data cleaning: Remove noise, redundant information, and possible erroneous input to ensure data consistency and accuracy.

[0020] Step 1.3: Knowledge representation: Divide and represent knowledge at different levels of granularity, such as documents, paragraphs, sentences, etc., and vectorize them using a pre-trained model.

[0021] h i =f(w i ; Θ),i=1,2,…,n

[0022] w irepresents the input text unit (such as a word or sentence), Θ represents the parameters of the model, f is a pre-trained model (such as BERT, GPT, etc.), and the generated h i It is the representation of text units in high-dimensional vector space.

[0023] Step 1.4: Entity Extraction. Use named entity recognition technology to identify key entities from the text. Entity recognition can be expressed as a sequence labeling problem. Through the conditional random field (CRF) or Transformer architecture, calculate the entity category probability of each word:

[0024]

[0025] Where X represents the input sequence, y t represents the label of the tth time step, and Θ is the parameter of the model.

[0026] Step 1.5: Relationship extraction. Use relationship extraction technology to identify the relationship between entities. The relationship extraction problem can be modeled as a classification problem. Suppose a given entity pair (e1, e2), its relationship is predicted by the following formula:

[0027] P(r|e1,e2,C; Θ)=softmax(W·g(e1,e2,C; Θ)+b)

[0028] where r is the relation category, C is the context, g is the relation encoding function, and W and b are classifier parameters.

[0029] Step 2: Knowledge generation and expansion.

[0030] Step 2.1: Expand descriptive knowledge. Describe the concept, entity, or event in detail, providing more context and details.

[0031] Step 2.2: Expand factual knowledge. Provide objective factual information, such as standards, statistics, historical events, etc.

[0032] Step 2.3: Expand common sense knowledge. Provide knowledge that is generally accepted by people, such as general rules in daily life, social norms, and cultural customs.

[0033] Step 2.4: Knowledge verification: Use the expanded knowledge to verify the answers generated by the model to ensure the accuracy of the answers.

[0034] Step 3: Knowledge retrieval and management.

[0035] Step 3.1: Knowledge retrieval: Combine keyword retrieval, semantic retrieval, intelligent semantic model question answering, and retrieval enhancement generation to achieve efficient knowledge retrieval.

[0036] Step 3.2: Knowledge management. When step 3.1 fails to retrieve appropriate knowledge, or the retrieved knowledge is incorrect, the knowledge can be updated to ensure normal subsequent use.

[0037] Step 4: Knowledge pattern discovery and adaptive retrieval.

[0038] Step 4.1: Multi-agent collaboration. Using multiple agents to work together, dynamically adjust the retrieval strategy and reasoning process according to the different intentions of the user's questions. Assume that the system has k agents A1, A2, ..., A k , then their collaborative work can be modeled as:

[0039]

[0040] Among them, w i represents the agent weight, R(A i ) represents agent A i Contribution to a specific task, O represents the output result.

[0041] Step 4.2: Adaptive retrieval strategy. Different retrieval strategies are used according to the different types of user questions. For simple and clear open-domain questions, the intelligent semantic model or the built-in knowledge of the intelligent agent is directly used to answer. For domain questions related to the local knowledge base, the retrieval augmentation generation (RAG) process is started and verified in combination with the knowledge graph. Before generating the final answer, multiple reasoning and loop verification are performed to ensure the correctness and rationality of the answer.

[0042] In addition, the present invention also provides a question-answering system integrating multi-granularity knowledge graphs and adaptive agents, which executes the aforementioned question-answering method integrating multi-granularity knowledge graphs and adaptive agents. The question-answering system includes a knowledge governance and extraction module, a knowledge generation and expansion module, a knowledge retrieval and management module, and an adaptive agent retrieval module.

[0043] The knowledge governance and extraction module includes data access unit, data cleaning unit, knowledge representation unit, entity extraction unit and relationship extraction unit;

[0044] The knowledge generation and expansion module includes descriptive knowledge units, factual knowledge units, common sense knowledge units, and knowledge verification units;

[0045] The knowledge retrieval and management module includes a knowledge retrieval unit and a knowledge management unit;

[0046] The adaptive agent retrieval module includes a multi-agent collaboration unit and an adaptive retrieval unit.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The question-answering method and system proposed in the present invention, which integrates multi-granularity knowledge graphs and adaptive agents, can more comprehensively understand user questions and retrieve more relevant knowledge from the knowledge base through the collaboration of multi-granularity knowledge graphs and adaptive agents, thereby significantly improving the accuracy of question-answering and reducing the generation of wrong answers. In addition, multi-granularity knowledge graphs can support more complex knowledge application requirements, such as reasoning and prediction, thereby enhancing the application capabilities of the question-answering system, enabling it to handle more complex question-answering tasks, such as multi-round dialogues, multi-task question-answering, etc. Comprehensively bring users a better user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 This is a flowchart of a question-answering method that integrates a multi-granularity knowledge graph and an adaptive agent in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0052] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0053] See also Figure 1 In one embodiment, the question-answering method integrating multi-granularity knowledge graph and adaptive agent of the present invention comprises the following steps:

[0054] Step 1: Knowledge governance and extraction.

[0055] Step 1.1: Data access. Support the access of multi-source heterogeneous data, including text, images, audio, etc., and uniformly process them through multimodal intelligent semantic models to convert non-text data into text form.

[0056] Step 1.2: Data cleaning: Remove noise, redundant information, and possible erroneous input to ensure data consistency and accuracy.

[0057] Step 1.3: Knowledge representation: Divide and represent knowledge at different levels of granularity, such as documents, paragraphs, sentences, etc., and vectorize them using a pre-trained model.

[0058] h i =f(w i ; Θ),i=1,2,…,n

[0059] w i represents the input text unit (such as a word or sentence), Θ represents the parameters of the model, f is a pre-trained model (such as BERT, GPT, etc.), and the generated h i It is the representation of text units in high-dimensional vector space.

[0060] Step 1.4: Entity Extraction. Use named entity recognition technology to identify key entities from the text. Entity recognition can be expressed as a sequence labeling problem. Through the conditional random field (CRF) or Transformer architecture, calculate the entity category probability of each word:

[0061]

[0062] Where X represents the input sequence, y t represents the label of the tth time step, and Θ is the parameter of the model.

[0063] Step 1.5: Relationship extraction. Use relationship extraction technology to identify the relationship between entities. The relationship extraction problem can be modeled as a classification problem. Suppose a given entity pair (e1, e2), its relationship is predicted by the following formula:

[0064] P(r|e1,e2,C; Θ)=softmax(W·g(e1,e2,C; Θ)+b)

[0065] where r is the relation category, C is the context, g is the relation encoding function, and W and b are classifier parameters.

[0066] Step 2: Knowledge generation and expansion.

[0067] Step 2.1: Expand descriptive knowledge. Describe the concept, entity, or event in detail, providing more context and details.

[0068] Step 2.2: Expand factual knowledge. Provide objective factual information, such as standards, statistics, historical events, etc.

[0069] Step 2.3: Expand common sense knowledge. Provide knowledge that is generally accepted by people, such as general rules in daily life, social norms, and cultural customs.

[0070] Step 2.4: Knowledge verification: Use the expanded knowledge to verify the answers generated by the model to ensure the accuracy of the answers.

[0071] Step 3: Knowledge retrieval and management.

[0072] Step 3.1: Knowledge retrieval: Combine keyword retrieval, semantic retrieval, intelligent semantic model question answering, and retrieval enhancement generation to achieve efficient knowledge retrieval.

[0073] Step 3.2: Knowledge management. When step 3.1 fails to retrieve appropriate knowledge, or the retrieved knowledge is incorrect, the knowledge can be updated to ensure normal subsequent use.

[0074] Step 4: Knowledge pattern discovery and adaptive retrieval.

[0075] Step 4.1: Multi-agent collaboration. Using multiple agents to work together, dynamically adjust the retrieval strategy and reasoning process according to the different intentions of the user's questions. Assume that the system has k agents A1, A2, ..., A k , then their collaborative work can be modeled as

[0076]

[0077] Among them, w i represents the agent weight, R(A i ) represents agent A i Contribution to a specific task, O represents the output result.

[0078] Step 4.2: Adaptive retrieval strategy. Different retrieval strategies are used according to the different types of user questions. For simple and clear open-domain questions, the intelligent semantic model or the built-in knowledge of the intelligent agent is directly used to answer. For domain questions related to the local knowledge base, the retrieval augmentation generation (RAG) process is started and verified in combination with the knowledge graph. Before generating the final answer, multiple reasoning and loop verification are performed to ensure the correctness and rationality of the answer.

[0079] In addition, the present invention also provides a question-answering system integrating multi-granularity knowledge graphs and adaptive agents, which executes the aforementioned question-answering method. The question-answering system includes a knowledge governance and extraction module, a knowledge generation and expansion module, a knowledge retrieval and management module, and an adaptive agent retrieval module.

[0080] The knowledge governance and extraction module includes a data access unit, a data cleaning unit, a knowledge representation unit, an entity extraction unit, and a relationship extraction unit. The data access unit supports the access of multi-source heterogeneous data, including text, images, audio, etc., and uniformly processes non-text data through a multimodal intelligent semantic model to convert non-text data into text form. The data cleaning unit is used to remove noise, redundant information, and possible erroneous input to ensure data consistency and accuracy. The knowledge representation unit divides and represents knowledge at different granularity levels, such as documents, paragraphs, sentences, etc., and uses a pre-trained model for vectorization. The entity extraction unit uses named entity recognition technology to identify key entities from the text. Entity recognition can be expressed as a sequence labeling problem. The entity category probability of each word is calculated through the conditional random field (CRF) or Transformer architecture. The relationship extraction unit uses relationship extraction technology to identify the relationship between entities.

[0081] The knowledge generation and expansion module includes descriptive knowledge units, factual knowledge units, common sense knowledge units, and knowledge verification units. Descriptive knowledge units are used to expand descriptive knowledge, describe concepts, entities, or events in detail, and provide more context and details. Factual knowledge units are used to expand factual knowledge and provide objective factual information, such as standards, statistics, historical events, etc. Common sense knowledge units are used to expand common sense knowledge and provide knowledge that is generally recognized by people, such as general laws, social norms, and cultural customs in daily life. The knowledge verification unit uses the expanded knowledge to verify the answers generated by the model to ensure the accuracy of the answers.

[0082] The knowledge retrieval and management module includes the knowledge retrieval unit and the knowledge management unit. The knowledge retrieval unit combines keyword retrieval, semantic retrieval, intelligent semantic model question and answer, and retrieval enhancement generation to achieve efficient knowledge retrieval. The knowledge management unit is used to update the knowledge when appropriate knowledge cannot be retrieved or the retrieved knowledge is incorrect, to ensure normal subsequent use.

[0083] The adaptive agent retrieval module includes a multi-agent collaborative unit and an adaptive retrieval unit. The multi-agent collaborative unit uses multiple agents to work together and dynamically adjusts the retrieval strategy and reasoning process according to the different intentions of the user's questions. The adaptive retrieval unit adopts different retrieval strategies according to the different types of user questions. For simple and clear open domain questions, it directly uses the intelligent semantic model or the built-in knowledge of the agent to answer. For domain questions related to the local knowledge base, the retrieval augmentation generation (RAG) process is started and verified in combination with the knowledge graph. Before generating the final answer, multiple reasoning and loop verification are performed to ensure the correctness and rationality of the answer.

[0084] Taking the medical scenario as an example but not limited to the medical scenario, the question-answering method and system proposed in the present invention that integrates multi-granularity knowledge graphs and adaptive intelligent agents will play an important role in providing doctors and patients with more efficient and accurate medical services.

[0085] First, the system will access various data in the medical field, such as medical records, literature, and drug instructions, and perform data cleaning and knowledge representation. Through named entity recognition technology, the system will identify key entities from the text, such as disease names, drug names, symptoms, etc., and use relationship extraction technology to identify the relationship between entities, such as "high blood pressure may cause heart disease."

[0086] Next, the system will use the expanded descriptive knowledge to describe the disease, symptoms, drugs, etc. in detail, and provide more medical knowledge, such as "the symptoms of hypertension include dizziness, headache, palpitations, etc." At the same time, the system will use the expanded factual knowledge to provide statistical data such as the incidence, mortality, and treatment plans of the disease, such as "the incidence of hypertension increases with age." In addition, the system will also use the expanded common sense knowledge to provide some medical common sense, such as "more exercise can enhance immunity," and the expanded logical rationality knowledge to ensure the coherence, consistency, and logic of the knowledge.

[0087] When a patient asks the system a question, such as "I have been feeling chest tightness and shortness of breath recently. What could be the cause? Which department of doctor should I see?", the system will retrieve relevant information from the knowledge base based on the patient's description and generate preliminary answers using technologies such as keyword retrieval, semantic retrieval, intelligent semantic model question and answer, and retrieval enhancement generation.

[0088] Then, the system will use the expanded knowledge to verify the preliminary answer to ensure the accuracy of the answer. For example, the system will determine whether the patient needs treatment for hypertension based on the patient's medical history, family history and other information. Finally, the system will adopt different retrieval strategies according to the type of patient's question. For example, for simple and clear questions, the system will directly use the built-in knowledge of the intelligent semantic model or intelligent agent to answer; for questions related to the local knowledge base, the retrieval enhancement generation (RAG) process will be started and verified in combination with the knowledge graph; before generating the final answer, multiple reasoning and loop verification will be performed to ensure the correctness and rationality of the answer.

[0089] Through the above steps, the question-answering system proposed by the present invention can provide doctors with fast and accurate knowledge retrieval services, helping them to better understand the patient's condition and formulate more effective treatment plans. At the same time, the system can also provide patients with more convenient and comprehensive medical consultation services, helping them to better understand their own health conditions and improve medical efficiency.

[0090] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0091] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this article is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of this article. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0092] It should be noted that the above specific embodiments are exemplary, and those skilled in the art can come up with various solutions inspired by the disclosure of the present invention, and these solutions also belong to the disclosure scope of the present invention and fall within the protection scope of the present invention. Those skilled in the art should understand that the present invention description and its drawings are illustrative and do not constitute a limitation of the claims. The protection scope of the present invention is defined by the claims and their equivalents.

Claims

1. A question-answering method integrating multi-granularity knowledge graph and adaptive agent, characterized in that: The method comprises: Step 1: Knowledge governance and extraction, which includes: Step 1.1: Data access: Support access to multi-source heterogeneous data, and then use a multimodal intelligent semantic model to uniformly process and convert non-text data into text form; multi-source heterogeneous data includes text, images, and audio; Step 1.2: Data cleaning: remove noise, redundant information and erroneous input to ensure data consistency and accuracy; Step 1.3: Knowledge representation: divide and represent knowledge at different levels of granularity and vectorize it using a pre-trained model: h i =f(w i ;Θ),i=1,2,…,n Among them, w i represents the input text unit, Θ represents the parameters of the model, f represents the pre-trained model, and the generated h i It is the representation of text units in high-dimensional vector space; Step 1.4: Entity extraction; Use named entity recognition technology to identify key entities from the text. Entity recognition can be expressed as a sequence labeling problem. Through the conditional random field (CRF) or Transformer architecture, the entity category probability of each word is calculated: Among them, X represents the input sequence, y t represents the label of the tth time step, Θ represents the parameters of the model; Step 1.5: Relationship extraction: Use relationship extraction technology to identify the relationship between entities; the relationship extraction problem can be modeled as a classification problem. Suppose a given entity pair (e1, e2), its relationship is predicted by the following formula: P(r|e1,e2,C; Θ)=softmax(W·g(e1,e2,C; Θ)+b) Where r is the relation category, C is the context, g is the relation encoding function, and W and b are classifier parameters; Step 2: Knowledge generation and expansion; Step 3: Knowledge retrieval and management; Step 4: Knowledge pattern discovery and adaptive retrieval, which includes: Step 4.1: Multi-agent collaboration: Use multiple agents to work together and dynamically adjust the retrieval strategy and reasoning process according to the different intentions of the user's questions. Assume that there are k agents A1, A2, ..., A k , then their collaborative work can be modeled as: Among them, w i represents the agent weight, R(A i ) represents agent A i Contribution to a specific task, O represents the output result; Step 4.2: Execute adaptive retrieval strategies; adopt different retrieval strategies according to the different types of user questions. For simple and clear open-domain questions, directly use the intelligent semantic model or the built-in knowledge of the intelligent agent to answer; for domain questions related to the local knowledge base, start the retrieval augmentation generation (RAG) process and verify it in combination with the knowledge graph; before generating the final answer, perform multiple reasoning and loop verification to ensure the correctness and rationality of the answer.

2. The method according to claim 1, characterized in that Step 2 specifically includes: Step 2.1: Expand descriptive knowledge; describe the concept, entity, or event in detail, providing more context and details; Step 2.2: Expand factual knowledge; provide objective factual information, such as standards, statistics, and historical events; Step 2.3: Expand common sense knowledge; provide knowledge that people generally agree on, such as general rules, social norms, and cultural customs in daily life; Step 2.4: Knowledge verification: Use the expanded knowledge to verify the answers generated by the model to ensure the accuracy of the answers.

3. The method according to claim 2, characterized in that In step 3, it also includes: Step 3.1: Knowledge retrieval: Combine keyword retrieval, semantic retrieval, intelligent semantic model question answering and retrieval enhancement generation technology to achieve efficient knowledge retrieval; Step 3.2: Knowledge management; when step 3.1 fails to retrieve appropriate knowledge, or the retrieved knowledge is incorrect, update the knowledge to ensure normal subsequent use.

4. A question-answering system integrating multi-granularity knowledge graph and adaptive agent, characterized in that: The question-answering system executes the method according to claim 1, and the question-answering system includes a knowledge management and extraction module, a knowledge generation and expansion module, a knowledge retrieval and management module, and an adaptive agent retrieval module, wherein: The knowledge governance and extraction module includes data access unit, data cleaning unit, knowledge representation unit, entity extraction unit and relationship extraction unit; The knowledge generation and expansion module includes descriptive knowledge units, factual knowledge units, common sense knowledge units, and knowledge verification units; The knowledge retrieval and management module includes a knowledge retrieval unit and a knowledge management unit; The adaptive agent retrieval module includes a multi-agent collaboration unit and an adaptive retrieval unit.

Images